INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
Lightweight Dual-View Feature Fusion for Hand-Object Interaction  
Recognition  
Houda Skhoun  
School of Artificial Intelligence Nanjing University of Information Science and Technology (NUIST)  
Nanjing, China  
DOI: https://doi.org/10.51583/IJLTEMAS.2026.150500198  
Received: 17 May 2026; Accepted: 22 May 2026; Published: 13 June 2026  
ABSTRACT  
Wrist-worn hand-object interaction (HOI) recognition is a critical capability for wearable rehabilitation  
systems, assistive technologies,augmented reality and human-computer interaction applications. Compared  
with fixed external cameras, wrist-worn wearable devices provide user-centered observations that are more  
suitable for continuous real-world interaction monitoring. However, many existing wrist-worn HOI  
recognition systems still face important challenges, including incomplete interaction representation caused  
by single-view observations, viewpoint ambiguity, self-occlusion and the high computational complexity  
of recent deep learning approaches. To address these limitations, this paper proposes a lightweight dual-  
view framework for hand-object interaction recognition using synchronized palm-view and back-view  
RGB images acquired from a wrist-worn dual-camera device. The proposed framework employs a shared  
MobileNetV2 backbone combined with multi-level feature extraction to jointly capture fine-grained spatial  
details and high-level semantic representations. To effectively integrate complementary information from  
different network depths, a view-specific adaptive fusion mechanism is introduced to dynamically balance  
intermediate and deep feature representations for each visual stream. The fused dual-view representation is  
subsequently used for interaction classification. Experimental evaluation under the Leave- One-Participant-  
Out (LOPO) cross-subject protocol demonstrates that the proposed framework achieves a mean accuracy  
of 82.36% and a mean F1-score of 81.65% while maintaining low computational complexity suitable for  
real-time wearable applications. Ablation studies further confirm the effectiveness of the proposed multi-  
level feature extraction and adaptive fusion strategy. The proposed approach provides an effective balance  
between recognition performance and computational efficiency for lightweight wearable HOI recognition  
systems.  
Keywords: Hand-object interaction recognition; Dual-view learning; Adaptive feature fusion; Lightweight  
deep learning; Wrist-worn systems  
INTRODUCTION  
Wrist-worn hand-object interaction (HOI) recognition has emerged as an important capability for  
wearable rehabilitation systems, assistive technologies, augmented reality (AR/VR), and human-robot  
collaboration [1–3]. Accurate recognition of object manipulation enables intelligent systems to better  
understand user intention and provide adaptive responses during real-world interaction scenarios.  
Recent advances in deep learning have significantly improved visual recognition performance by  
enabling automatic extraction of hierarchical feature representations directly from image data, overcoming many  
limitations associated with traditional handcrafted feature approaches [4–8].  
Among existing acquisition settings, wrist-worn wearable systems have attracted increasing attention due  
to their portability, non-intrusive design and suitability for continuous real-world interaction monitoring.  
Compared with fixed external cameras, wrist-worn devices provide observations that remain closely  
aligned with the user’s hand movements during object manipulation while reducing dependency on  
environmental setup and camera placement. Their compact and wearable nature makes them particularly  
suitable for assistive technologies, rehabilitation systems and real- time human-computer interaction  
Page 2478  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
applications.  
Existing hand-object interaction recognition methods have explored various sensing configurations,  
including external cameras, egocentric wearable cameras, RGB-D systems and multi-modal sensing  
approaches [9–13]. Despite recent progress, robust HOI recognition remains challenging in wrist-worn  
settings due to several factors. First, many approaches rely on single-view observations, which may fail to  
capture complete interaction information, including fine-grained finger articulation, object contact regions  
and global hand posture. Second, viewpoint variation, illumination changes and inter-subject differences in  
hand shape and manipulation style further degrade recognition performance [2, 14]. Finally, many recent  
methods depend on computationally expensive architectures or temporal modeling strategies, limiting their  
suitability for lightweight real-time wearable applications.  
To address these limitations, this paper proposes a lightweight dual-view framework for hand-object  
interaction recognition using synchronized palm-view and back-view observations acquired from a wrist-  
worn dual-camera device. The proposed acquisition setup provides complementary visual information from  
two synchronized viewpoints, where the palm-view stream captures detailed finger articulation and object  
contact regions, while the back-view stream provides global hand posture and movement configuration. By  
combining both visual streams, the proposed framework improves robustness against self-occlusion and  
viewpoint ambiguity during fine- grained interaction analysis.  
Based on this acquisition setting, a lightweight multi-level deep learning framework is developed using a shared  
MobileNetV2 backbone combined with adaptive feature fusion. Intermediate and high-level feature  
representations are extracted from different network depths to capture complementary spatial and semantic  
information. An adaptive gating mechanism is then employed to dynamically balance the contribution of multi-  
level features for each visual stream before dual-view fusion and interaction classification. The proposed  
framework is designed to maintain low computational complexity while achieving strong recognition  
performance suitable for wearable real-time applications.  
The proposed framework is evaluated using a subject-independent Leave-One-Participant-Out (LOPO)  
protocol on a dual-view hand-object interaction dataset collected using the proposed wrist-worn acquisition  
system. Experimental results demonstrate that the proposed framework achieves strong recognition  
performance while maintaining low computational cost under cross-subject evaluation conditions.  
The main contributions of this paper are summarized as follows:  
1.A lightweight dual-view framework is proposed for hand-object interaction recognition using synchronized  
palm-view and back-view observations acquired from a wrist-worn dual-camera device.  
2.A multi-level feature extraction strategy is developed to integrate intermediate and deep representations for fine-  
grained interaction analysis.  
3.An adaptive view-specific fusion mechanism is proposed to dynamically combine complementary multi-level  
representations from the two visual streams.  
4.Extensive experiments under the LOPO cross-subject evaluation protocol demonstrate the effectiveness,  
robustness and computational efficiency of the proposed frame-work for wearable real-time HOI  
recognition.  
RELATED WORK  
Hand-Object Interaction Recognition  
Hand-object interaction (HOI) recognition has attracted significant attention in computer vision due to its  
applications in human-computer interaction, rehabilitation systems, robotics and wearable intelligent  
devices [1–3]. Compared with isolated hand gesture recognition, HOI recognition is more challenging  
Page 2479  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
because interaction understanding depends on both hand configuration and object-related contextual  
information. In practical scenarios, severe occlusions, viewpoint variations, illumination changes and inter-  
subject differences further complicate robust recognition [2, 14].  
Early HOI recognition methods mainly relied on handcrafted features and conventional machine learning  
techniques. Traditional approaches focused on hand segmentation object tracking, and geometric modeling  
using RGB or RGB-D inputs [15–18]. Other works incorporated contextual information such as grasp type  
and object attributes to better characterize manipulation actions [11, 19]. Although these methods  
established important foundations for interaction analysis, their dependence on handcrafted representations  
and multi-stage processing pipelines limited their generalization capability in unconstrained environments.  
Recent advances in deep learning have signifi-cantly improved HOI recognition by enabling end-to-end  
learning of discriminative visual representations [4–6]. CNN-based approaches have been widely adopted  
for egocentric interaction under-standing, object localization and hand-object relationship modeling  
[3,20,21]. In addition, several studies proposed multitask frameworks integrating hand pose estimation,  
object detection and interaction recognition within unified architectures [12, 22, 23]. Temporal modeling  
techniques based on recurrent networks, 3D convolutional networks and ConvLSTM architectures further  
improved dynamic interaction analysis [9, 24–29]. Multi-modal approaches combining RGB, depth and  
skeleton information have also been explored to improve robustness under challenging conditions [10, 30,  
31].  
Despite these advances, many existing HOI recognition methods rely on large annotated datasets, temporal video  
modeling or computationally expensive architectures, which may limit their applicability in lightweight wearable  
systems and real-time environments [3, 12, 23]. Furthermore, single-view observations often provide incomplete  
interaction representations under occlusion and viewpoint variation, motivating the exploration of multi-view  
learning strategies.  
Multi-View Learning and Feature Fusion  
Multi-view learning has been widely investigated to improve recognition robustness under viewpoint  
variation and partial occlusion [10]. By combining complementary observations from multiple view-  
points, multi-view systems can capture more complete spatial information and reduce ambiguity during  
visual recognition tasks. Previous studies demonstrated that multi-view observations improve gesture and  
action recognition performance by exploiting complementary visual cues across different perspectives [33–  
35].  
To integrate information from multiple view-points, several fusion strategies have been proposed. Early fusion  
combines visual information before feature extraction, enabling the network to directly learn correlations  
between views [36, 37]. Feature-level fusion merges intermediate representations extracted independently from  
each view- point while preserving discriminative characteristics from individual views [38, 39]. Late fusion  
combines predictions generated from separate visual streams [36, 40]. More recent approaches employ attention-  
based or adaptive fusion mechanisms to dynamically emphasize the most informative representations according  
to the input characteristics [41–43]. These methods demonstrated improved robustness in recognition tasks  
affected by viewpoint ambiguity and occlusion.  
However, many existing multi-view approaches rely on computationally intensive architectures or dense  
temporal processing. In addition, several methods primarily focus on global representations without  
explicitly exploiting complementary information from different feature levels.  
Page 2480  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
Figure 1: (a) Custom-designed wearable dual-camera device showing the positioning of the back and palm  
cameras at approximately 2 cm from the wrist. (b) Examples of captured images from the two viewpoints,  
illustrating the back view and palm view with a 160° wide-angle field of view.  
Multi-Level Feature Learning  
Multi-level feature learning has become an important component of modern deep visual recognition  
systems. CNNs naturally learn hierarchical representations, where shallow and intermediate layers capture  
local spatial patterns, while deeper layers encode high-level semantic information. Residual learning and  
feature pyramid strategies demonstrated that combining features from multiple network depths improves  
robustness in complex visual recognition tasks [44, 45].  
For HOI recognition, multi-level feature integration is particularly important because subtle variations in  
finger articulation, grasp configuration and object contact regions often determine the interaction category.  
However, many existing approaches mainly rely on deep semantic representations while ignoring  
intermediate spatial features that may contain important fine-grained interaction cues. Moreover, several  
methods continue to depend on single-view observations or computationally expensive temporal  
architectures, limiting their suitability for lightweight wearable applications and real-time deployment.  
Another important challenge concerns subject-independent generalization. Variations in hand shape,  
manipulation style and interaction execution across users can significantly affect visual appearance and  
reduce recognition robustness. Consequently, there remains a need for efficient HOI recognition  
frameworks capable of integrating complementary multi-view information while maintaining low  
computational complexity and strong cross-subject generalization capability.  
To address these limitations, this work proposes a lightweight dual-view multi-level feature fusion  
framework for hand-object interaction recognition using synchronized palm-view and back-view ob-  
servations. The proposed framework combines multi-level feature extraction with adaptive view- specific  
fusion to effectively exploit complementary visual information while maintaining computational efficiency  
suitable for wearable real-time applications.  
Page 2481  
www.rsisinternational.org  
 
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
Figure 2: Overview of the proposed dual-view multi-level feature fusion framework for hand object  
interaction recognition. The framework integrates dual-view inputs, multi-level feature extraction, adaptive  
gating fusion and final classification  
METHODOLOGY  
Data Acquisition  
The proposed framework employs a lightweight wrist-worn dual-camera device designed to capture  
complementary visual information during hand-object interactions. As illustrated in Figure 1, two  
synchronized wide-angle RGB cameras with an approximate 160◦ field of view are mounted on a wearable  
wrist support to simultaneously acquire palm-view and back-view observations of the same interaction  
instance. The back-view camera captures global hand posture and interaction context, whereas the palm-  
view camera provides detailed information related to finger articulation, contact regions and manipulated  
objects. The cameras are positioned approximately 2 cm from the wrist surface to maximize viewpoint  
complementarity while maintaining a compact wearable configuration.  
Let Ip and Ib denote the palm-view and back- view RGB inputs, respectively. Each synchro- nized  
image pair represents the same interaction captured from complementary viewpoints. This dual-view  
representation enables the framework to exploit both fine-grained local interaction cues and global  
contextual information while reducing ambiguities caused by self-occlusion and view- point variation. For  
compatibility with the MobileNetV2 backbone, all input images are resized to 224×224 pixels. During  
training, data augmentation techniques including random resized cropping, color jittering, affine  
transformations, Gaus- sian blurring and random grayscale conversion are applied to improve robustness  
and reduce overfitting. Since the two images form a synchronized pair, identical augmentation parameters  
are applied to both views to preserve spatial correspondence. The processed images are finally normal- ized  
using ImageNet mean and standard deviation values before being passed to the network.  
Network Overview  
The proposed framework performs hand-object interaction (HOI) recognition using synchronized palm-  
view and back-view observations. As illustrated in Figure 2, the framework consists of dual visual  
streams, multi-level feature extraction, adaptive feature fusion and a lightweight classification module.  
Let Ip  
∈
R3×H×W and Ib  
∈
R3×H×W denote the palm-view and back-view RGB inputs, respectively. Both  
visual streams are processed using a shared MobileNetV2 backbone Φ(·) to extract hierarchical feature  
representations Fp and Fb. To preserve both local spatial details and high-level semantic information, multi-  
level features are extracted from intermediate and deep network layers. An adaptive gating module then  
generates compact view-specific representations vpalm and vback for the palm-view and back-view streams,  
respectively.  
The resulting representations are subsequently fused to generate the final interaction  
representation v, which is passed to the classification module for HOI prediction. The proposed framework  
effectively exploits complementary multi-view information while maintaining computational efficiency  
adapted to wearable real-time applications.  
Page 2482  
www.rsisinternational.org  
 
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
Figure 3: Multi-level feature extraction and token generation from the MobileNetV2 backbone  
Multi-Level Feature Extraction  
To obtain discriminative yet efficient visual representations, the proposed framework employs Mo-  
bileNetV2 as a shared lightweight backbone for both palm-view and back-view streams. MobileNetV2 is  
selected due to its favorable trade-off between recognition performance and computational efficiency, making  
it compatible with wearable real-time applications.  
Given the synchronized inputs Ip and Ib, the shared backbone Φ(·) extracts hierarchical feature  
representations for each visual stream:  
Fp = Φ(Ip),  
Fb = Φ(Ib)  
(1)  
Instead of relying only on the final deep representation, the proposed framework extracts features from  
multiple network depths to preserve both local spatial details and high-level semantic information. As  
illustrated in Figure 3, intermediate feature maps capture fine-grained interaction patterns related to  
finger articulation, object boundaries and contact regions, whereas deeper feature maps encode higher-  
level interaction semantics. For each visual stream, the backbone produces an intermediate feature  
map Fm and a deep feature map Fd. Global average pooling is then applied to obtain compact feature  
vectors:  
zm = GAP(Fm),  
zd = GAP(Fd)  
(2)  
Since the extracted representations have different dimensions, lightweight linear projection layers are used  
to map them into a unified embedding space:  
vmid  
=
Wmzm  
+
bm,  
vdeep  
=
Wdzd  
+
bd  
(3)  
where Wm and Wd denote learnable projection matrices and bm and bd are bias terms. The projected  
representations vmid and vdeep are subsequently used as inputs to the adaptive fusion module.  
Adaptive Feature Fusion  
The discriminative importance of intermediate and deep feature representations may vary depending on the  
interaction type, viewpoint conditions and object appearance. To dynamically balance the contribution of  
different representation levels, the proposed framework employs an adaptive gating module for each visual  
stream, as illustrated in Figure 4. Given the projected intermediate and deep representations vmid and vdeep  
,
adaptive gating weights are computed as:  
Page 2483  
www.rsisinternational.org  
 
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
ω
=
MLP([vmid; vdeep])  
(4)  
where [;] denotes feature concatenation and ω [ωmid; ωdeep] represents the learned fusion weights  
=
associated with the intermediate and deep representations, respectively. To obtain normalized adaptive  
weights, a softmax operation is applied:  
[αmid, αdeep]  
=
Softmax(ω)  
(5)  
The final view-specific representation vview is then computed as a weighted combination of the two feature levels:  
view  
v
=
α
⊙
v
+
α
⊙
v
deep  
(6)  
mid  
mid  
deep  
Figure 4: View-specific adaptive gating mechanism for combining mid-level and deep-level feature  
representations  
where  
⊙
denotes element-wise multiplication. This adaptive formulation enables the framework to  
dynamically emphasize either fine-grained spatial information or high-level semantic representations  
according to the characteristics of the interaction. After adaptive fusion, the palm-view and back-view  
representations are concatenated to generate the final interaction representation:  
v = [vpalm; vback]  
(7)  
By adaptively balancing multi-level representations for each visual stream, the proposed fusion strategy  
effectively exploits complementary information from both viewpoints while maintaining low  
computational complexity appropriate for lightweight wearable applications.  
Classification  
The fused representation v is passed to a lightweight classification module composed of two fully connected  
layers with GELU activation and dropout regularization. The classifier transforms the fused representation  
into the final hand-object interaction prediction scores. The network is trained end-to-end using the cross-  
entropy loss function. The resulting output scores correspond to the predicted hand-object interaction  
categories.  
EXPERIMENTS AND RESULTS  
Experimental Setup  
Dataset Description  
The experiments are conducted on a dual-view hand-object interaction dataset collected using the proposed  
wrist-worn acquisition system. Each interaction instance consists of synchronized palm-view and back-  
Page 2484  
www.rsisinternational.org  
 
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
view RGB image pairs captured simultaneously from complementary viewpoints. Example interaction  
samples from the proposed dataset are illustrated in Figure 5.  
The dataset contains 13 hand-object  
interaction categories involving four commonly manipulated objects: pen, mouse, book and phone. The  
interaction types include grasp, hold, pinch and support actions. Data were collected from 10 participants  
under varying hand configurations and interaction styles to ensure cross-subject diversity. Overall, the  
dataset contains 26,894 synchronized interaction samples with a near-uniform class distribution, as  
summarized in Table 1. The diversity of subjects, viewpoints and interaction patterns provides a challenging  
evaluation scenario for cross-subject hand-object interaction recognition under varying occlusion and  
viewpoint conditions.  
Table 1: Class-wise distribution of interaction samples in the proposed dual-view hand object interaction  
dataset  
Class ID  
C01  
Interaction  
Grasp Pen  
#Samples  
2092  
C02  
Hold Book  
Hold Mouse  
Hold Pen  
2052  
C03  
2059  
C04  
2059  
C05  
Hold Phone  
Pinch Book  
Pinch Mouse  
Pinch Pen  
2051  
C06  
2071  
C07  
2095  
C08  
2077  
C09  
Pinch Phone  
Support Book  
Support Mouse  
Support Pen  
Support Phone  
—
2060  
C10  
2052  
C11  
2059  
C12  
2110  
C13  
2057  
Total  
26,894  
Evaluation Protocol  
To evaluate the cross-subject generalization capability of the proposed framework, a Leave-One-  
Participant-Out (LOPO) cross-validation protocol is adopted. The dataset contains interaction samples  
collected from 10 participants. In each evaluation fold, the samples from one participant are used for testing,  
while the remaining participants are used for training. This process is repeated across all participants and  
the final performance is reported as the average across the 10 folds. The LOPO protocol ensures strict  
subject-independent evaluation by testing the model on unseen users, providing a reliable assessment of the  
framework under varying hand shapes, interaction styles and manipulation patterns.  
Page 2485  
www.rsisinternational.org  
 
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
Figure 5: Each interaction instance is represented by synchronized palm and back views, highlighting  
complementary visual information for different grasp types and manipulated objects  
Implementation Details  
The proposed framework is implemented using the PyTorch deep learning library and trained using GPU  
acceleration. The network is trained for 20 epochs with a batch size of 32 using the AdamW optimizer. The  
initial learning rate is set to 3 × 10−5, with a weight decay of 5 × 10−5. A cosine annealing learning rate  
scheduler is employed to gradually reduce the learning rate during training. Cross-entropy loss with label  
smoothing 0.1 is adopted to improve generalization.  
During training, synchronized data augmentation is applied to both visual streams. The images are  
normalized using the ImageNet mean and standard deviation values. For evaluation, images are resized to  
224×224, normalized using ImageNet statistics and tested without augmentation. Experiments are  
Page 2486  
www.rsisinternational.org  
 
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
conducted under the Leave-One-Participant-Out (LOPO) cross-validation protocol, where each participant  
is used once as the test subject while the remaining participants are used for training.  
Table 2: Cross-subject performance of the proposed method under the Leave-One-Participant-Out (LOPO)  
protocol  
Fold  
Acc (%)  
89.95  
F1 (%)  
90.57  
89.04  
81.31  
81.91  
92.57  
89.21  
77.65  
88.63  
88.86  
88.58  
Precision (%)  
89.89  
Recall (%)  
89.94  
1
2
87.26  
87.38  
87.00  
3
71.50  
72.02  
70.63  
4
76.49  
76.75  
76.71  
5
91.28  
91.57  
91.06  
6
86.73  
86.73  
86.32  
7
63.09  
63.65  
59.42  
8
86.99  
86.71  
86.20  
9
84.63  
84.85  
83.61  
10  
85.68  
85.77  
85.58  
Mean  
±std  
82.36  
±
8.61  
86.83  
±
4.55  
82.53  
±
8.43  
81.65  
±
9.42  
Evaluation Metrics  
The performance of the proposed framework is evaluated using accuracy, macro-averaged precision, recall  
and F1-score. Accuracy measures the overall proportion of correctly classified interaction samples, while  
precision, recall and F1-score provide a more comprehensive evaluation of the recognition performance  
across all interaction categories. The macro-average strategy computes each metric independently for every  
class and then averages the results, ensuring equal contribution from all categories regardless of sample  
frequency. This evaluation protocol provides a balanced assessment of the framework under the cross-  
subject LOPO setting. In addition, confusion matrix analysis is performed to investigate class-wise  
recognition behavior and identify common misclassification patterns between visually similar hand-object  
interaction categories.  
Quantitative Results  
LOPO Cross-Subject Performance  
To evaluate the cross-subject generalization capability of the proposed framework, experiments are  
conducted using the Leave-One-Participant-Out (LOPO) evaluation protocol described in the previous  
section. The model is trained using data from nine participants and evaluated on the remaining unseen  
participant across all 10 folds.  
The quantitative results are summarized in Table 2. The proposed framework achieves a mean accuracy of  
82.36% and a mean F1-score of 81.65%, demonstrating strong recognition performance under the subject-  
independent evaluation setting. The framework also achieves high precision and recall values, indicating  
Page 2487  
www.rsisinternational.org  
 
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
stable classification performance across different interaction categories and participants.  
The results demonstrate that the proposed dual-view framework effectively captures complementary  
interaction information from the palm and back views while maintaining robustness to variations in hand  
appearance, manipulation style and viewpoint conditions. Despite the challenging cross-subject evaluation  
setting, the framework maintains consistent recognition performance across most participants, highlighting  
the effectiveness of the proposed multi-level feature extraction and adaptive fusion strategy.  
Comparison with Baseline Methods  
To evaluate the effectiveness of the proposed framework, comparisons are conducted with several widely  
used lightweight convolutional neural network architectures, including MobileNetV2, MobileNetV3,  
ShuffleNetV2, Ghost-Net, EfficientNet-B0, MobileViT, SqueezeNet and MnasNet. All models are  
evaluated under the same LOPO cross-subject protocol using identical training settings and preprocessing  
strategies to ensure a fair comparison.  
Table 3: Performance comparison between the proposed method and baseline convolutional neural  
network architectures in terms of mean accuracy (mAcc), mean F1-score (mF1), mean precision, mean  
recall and test loss under the LOPO cross-subject evaluation protocol  
Method  
mAcc (%) mF1 (%)  
mPrec (%) mRec (%)  
Loss  
MobileNetV2  
MobileNetV3-Small  
MobileNetV3-Large  
ShuffleNetV2  
GhostNet  
80.60 ± 10.60  
68.74 ± 12.27  
76.33 ± 11.75  
75.04 ± 12.63  
73.88 ± 16.95  
77.75 ± 13.18  
77.07 ± 12.29  
75.35 ± 14.72  
77.69 ± 10.77  
79.58 ± 11.80  
66.73 ± 13.46  
75.06 ± 12.69  
73.57 ± 14.11  
71.99 ± 18.83  
76.18 ± 14.72  
75.73 ± 14.12  
73.86 ± 16.24  
76.83 ± 11.80  
85.62 ± 6.16  
74.71 ± 9.06  
82.63 ± 8.14  
79.58 ± 11.30  
78.72 ± 13.32  
83.08 ± 9.35  
82.81 ± 6.61  
78.93 ± 12.50  
83.33 ± 7.31  
80.82 ± 10.43  
68.99 ± 12.26  
76.57 ± 11.54  
75.24 ± 12.43  
74.21 ± 16.81  
77.97 ± 13.05  
77.40 ± 12.12  
75.67 ± 14.57  
77.94 ± 10.74  
1.03 ± 0.27  
1.33 ± 0.28  
1.17 ± 0.32  
1.16 ± 0.32  
1.22 ± 0.44  
1.11 ± 0.35  
1.14 ± 0.34  
1.20 ± 0.37  
1.17 ± 0.26  
EfficientNet-B0  
MobileViT  
SqueezeNet  
MnasNet  
82.36  
±
8.61  
81.65  
±
9.42  
86.83  
±
4.55  
82.53  
±
8.43  
1.01  
±
0.27  
Proposed Method  
Since the proposed framework operates on synchronized dual-view inputs, all baseline architectures are  
adapted to the same two-view setting. Feature representations extracted from the palm-view and back-view  
streams are combined using feature concatenation before classification.  
The quantitative comparison results are summarized in Table 3. The proposed framework achieves the best  
overall performance among the evaluated lightweight architectures, obtaining a mean accuracy of 82.36%,  
a mean F1-score of 81.65% and a mean precision of 86.83%. In comparison, MobileNetV2 achieves  
80.60% accuracy, while EfficientNet-B0 and MobileViT achieve 77.75% and 77.07%, respectively. The  
proposed framework also achieves the lowest test loss among most evaluated methods.  
These results demonstrate that the proposed multi-level feature extraction and adaptive fusion strategy  
effectively improve recognition performance under the cross-subject evaluation setting. Despite its  
lightweight design, the proposed framework achieves superior recognition performance.  
Page 2488  
www.rsisinternational.org  
   
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
Computational Complexity Analysis  
To evaluate the efficiency of the proposed framework, a computational complexity comparison is  
conducted using the number of parameters, FLOPs and inference latency. The results are summarized in  
Table 4. The proposed framework requires only 2.53M parameters and 0.65 GFLOPs, remaining  
comparable to lightweight architectures such as MobileNetV2 and Mnas-Net while being substantially less  
complex than EfficientNet-B0 and MobileViT. The framework achieves an average inference latency of  
16.94 ms, supporting real-time execution for wearable hand-object interaction recognition. Although  
SqueezeNet achieves lower latency due to its compact architecture, its recognition performance is  
significantly lower. In contrast, models such as GhostNet and EfficientNet-B0 exhibit higher computational  
cost and inference latency without providing superior recognition performance. These results indicate that  
the proposed framework provides an effective trade-off between recognition accuracy and computational  
efficiency for lightweight dual-view HOI recognition.  
Table 4: Computational complexity and efficiency comparison of the evaluated models in terms of  
parameters, FLOPs and inference latency  
Method  
Params  
2.88  
FLOPs  
0.65  
Latency  
16.19 ± 2.00  
15.21 ± 0.90  
19.24 ± 3.20  
23.91 ± 1.74  
32.51 ± 1.89  
32.06 ± 2.09  
29.55 ± 3.73  
6.16 ± 0.40  
15.45 ± 0.65  
MobileNetV2  
MobileNetV3-Small  
MobileNetV3-Large  
ShuffleNetV2  
GhostNet  
1.23  
0.12  
3.47  
0.47  
1.78  
0.30  
4.56  
0.31  
EfficientNet-B0  
MobileViT  
4.67  
0.77  
5.27  
2.84  
SqueezeNet  
0.99  
0.53  
MnasNet  
3.76  
0.67  
16.94  
±
1.30  
Proposed Method  
2.53  
0.65  
Ablation Study  
To evaluate the contribution of the different components of the proposed framework, an ablation study is  
conducted under the LOPO cross-subject evaluation protocol. The study investigates the impact of multi-  
level feature extraction and different feature fusion strategies while maintaining the same backbone  
architecture and training configuration across all variants.  
The quantitative results are summarized in Table 5. Using only deep features achieves a mean accuracy of  
79.64% and an F1-score of 78.87%, indicating that high-level semantic representations alone are insufficient to  
fully capture fine-grained interaction patterns. Incorporating both intermediate and deep features improves  
recognition performance, with concatenation and summation strategies achieving 80.36% and 81.63% accuracy,  
respectively.  
Page 2489  
www.rsisinternational.org  
 
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
Table 5: Ablation study evaluating the impact of multi-level feature extraction and fusion strategies on  
cross-subject recognition performance under the LOPO protocol  
Variant  
mAccuracy (%) mF1-score (%) Params (M) Latency (ms)  
Deep only  
79.64  
80.36  
81.63  
79.82  
82.36  
78.87  
79.92  
80.89  
78.93  
81.65  
2.46  
2.50  
2.47  
2.50  
2.53  
16.95 ± 1.99  
17.58 ± 2.44  
17.31 ± 2.59  
16.30 ± 1.46  
Mid+Deep+Concat  
Mid+Deep+Sum  
Mid+Deep+Shared Gate  
16.94  
±
1.30  
Mid+Deep+Separate  
Gate(Proposed)  
The shared adaptive gating mechanism achieves 79.82% accuracy, suggesting that a single gating function  
is insufficient to model the distinct characteristics of the palm-view and back-view streams. In contrast, the  
proposed separate gating mechanism achieves the best overall performance with 82.36% accuracy and an  
F1-score of 81.65%, while maintaining low computational complexity and real-time inference latency.  
These results demonstrate that the proposed view-specific adaptive fusion strategy enables more effective  
integration of complementary multi-level representations for dual-view hand-object interaction recognition.  
Qualitative Analysis  
Confusion Matrix Analysis  
To further investigate the classification behavior of the proposed framework, the aggregated confusion  
matrix obtained from the 10-fold LOPO evaluation is presented in Figure 6. Most predictions are strongly  
concentrated along the main diagonal, indicating effective recognition performance across the majority of  
interaction categories.  
Figure 6: Aggregated confusion matrix obtained from the 10-fold Leave-One-Participant-Out (LOPO)  
evaluation, illustrating the classification performance of the proposed method across the 13 hand object  
Page 2490  
www.rsisinternational.org  
 
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
interaction categories, with correct predictions concentrated along the main diagonal  
Several categories exhibit particularly high recognition accuracy. For example, support mouse (1924),  
support phone (1968) and support book (1866) achieve a large number of correct predictions. Similarly,  
strong recognition performance is observed for hold mouse (1818), hold pen (1787) and hold phone (1746).  
These interactions are generally characterized by more stable hand configurations and clearer object  
visibility.  
Despite the overall strong performance, several informative misclassification patterns are observed. In  
particular, confusion occurs among visually similar pinch-related interactions, especially between pinch  
pen and pinch phone (475 samples), as well as between pinch mouse and pinch phone (370 samples).  
Additional confusion is observed be- tween hold and pinch interactions involving the same object category,  
such as pinch book misclassified as hold book (205 samples). These patterns reflect the fine-grained nature  
of the task and the difficulty of distinguishing subtle differences in finger articulation and object appearance  
under partial occlusion.  
Overall, the confusion matrix demonstrates that the proposed framework effectively distinguishes most  
interaction categories while maintaining ro- bustness under cross-subject and multi-view evaluation  
conditions.  
Feature Representation Analysis  
To further analyze the learned feature representations, a t-SNE visualization is conducted using feature  
embeddings extracted from the proposed model at different training stages under the LOPO evaluation  
protocol. The visualization includes samples from all 13 interaction categories and is illustrated in  
Figure 7.  
Figure 7: t-SNE visualization of the learned feature representations at different training stages: Epoch 1,  
Epoch 10 and Epoch 20 for a representative subject under the LOPO evaluation protocol  
At Epoch 1, the feature representations already exhibit an initial level of class separation, indicating that the  
framework quickly captures meaningful interaction patterns from the dual-view inputs. However, several clusters  
remain relatively dispersed with partial overlap between visually similar interaction categories.  
As training progresses, the feature distributions become progressively more structured and compact. At Epoch  
10, clearer separation between interaction categories is observed, indicating improved discriminative  
representation learning. By Epoch 20, the feature embeddings form well-defined and compact clusters with  
reduced inter-class overlap, demonstrating that the proposed framework learns semantically meaningful  
representations for fine-grained hand-object interaction recognition.  
Page 2491  
www.rsisinternational.org  
 
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
Scalability Across Backbone Architectures  
To further investigate the scalability of the proposed framework, additional experiments are conducted by  
replacing the lightweight MobileNetV2 backbone with the higher-capacity ConvNeXtV2-Tiny architecture  
while preserving the proposed multi-level feature extraction and adaptive fusion strategy. In addition, the  
standalone ConvNeXtV2-Tiny backbone is evaluated under the same LOPO protocol to analyze the  
contribution of the proposed framework beyond the backbone architecture itself.  
The results are summarized in Table 6. The standalone ConvNeXtV2-Tiny backbone achieves a mean  
accuracy of 86.95% and a mean F1-score of 86.25%, demonstrating strong capability for fine-grained hand-  
object interaction recognition. When integrated into the proposed framework, the performance further  
improves to 87.84% mean accuracy and 86.98% mean F1-score, highlighting the effectiveness of the  
proposed multi-level feature extraction and adaptive fusion strategy in exploiting complementary dual-view  
information. Compared with the lightweight MobileNetV2-based configuration, the ConvNeXtV2-Tiny-  
based framework achieves higher recognition performance, but with increased computational complexity,  
as the number of parameters rises from 2.53M to 28.07M. These results demonstrate that the proposed  
framework is flexible and scalable across different backbone architectures, enabling adaptation to various  
computational and performance requirements.  
CONCLUSION  
This paper presented a lightweight dual-view framework for hand-object interaction recognition using  
synchronized palm-view and back-view observations acquired from a wrist-worn device. The proposed  
approach combines multi-level feature extraction and adaptive feature fusion to effectively capture both  
fine-grained interaction details and high-level semantic information from complementary viewpoints.  
To improve representation learning, the frame- work integrates intermediate and deep features extracted  
from a shared MobileNetV2 backbone and employs a view-specific adaptive gating mechanism to  
dynamically balance their contributions. Experimental results under the Leave-One-Participant-Out  
(LOPO) cross-subject evaluation protocol demonstrate that the proposed framework achieves strong  
recognition performance while maintaining low computational complexity suitable for real-time wearable  
applications.  
The ablation study confirms the effectiveness of the proposed multi-level fusion strategy and the importance  
of the separate adaptive gating mechanism for palm-view and back-view streams. Additional experiments  
further demonstrate the scalability of the framework across different backbone architectures.  
Overall, the proposed framework provides an effective and computationally efficient solution for dual-view  
hand-object interaction recognition and shows strong potential for future wearable intelligent systems,  
human-computer interaction and assistive technology applications.  
Table 6: Performance comparison of the standalone ConvNeXtV2-Tiny backbone and the proposed dual-  
view framework with different backbone architectures under the LOPO protocol  
Configuration  
mAcc (%)  
86.95  
mF1 (%)  
86.25  
Params (M)  
28.26  
ConvNeXtV2-Tiny  
82.36  
81.65  
2.53  
Proposed (MobileNetV2 backbone)  
Proposed (ConvNeXtV2 backbone)  
87.84  
86.98  
28.07  
REFERENCES  
1. Ohn-Bar E, Trivedi M M. Hand gesture recognition in real time for automotive in- terfaces[J].  
Page 2492  
www.rsisinternational.org  
   
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
IEEE Transactions on Intelligent Transportation Systems, 2014, 15(6): 2368–2377.  
2. Cheng H, Yang L, Liu Z. Survey on 3D hand gesture recognition[J]. IEEE Transactions  
on Circuits and Systems for Video Technology, 2015, 26(9): 1659–1673.  
3. Fan H, Zhuo T, Yu X, et al. Understanding atomic hand-object interaction with human  
intention[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2021, 32(1):  
275–285.  
4. Chung H Y, Chung  
Y
L, Tsai W F. An effi- cient hand gesture recognition system based on  
deep CNN[C]// IEEE International Con- ference on Industrial Technology. IEEE, 2019:  
853–858.  
5. Lin H I, Hsu M H, Chen W K. Human hand gesture recognition using a convolution neu- ral  
network[C]// IEEE International Confer- ence on Automation Science and Engineer- ing.  
IEEE, 2014: 1038–1043.  
6. Li G, Tang H, Sun Y, et al. Hand gesture recognition based on convolution neural net- work[J].  
Cluster Computing, 2019, 22(S2): 2719–2729.  
7. Ozcan T, Basturk A. Transfer learning-based convolutional neural networks with heuristic  
optimization[J]. Neural Computing and Ap- plications, 2019, 31(12): 8955–8970.  
8. Sahoo  
J
P, Prakash A J, Pl-awiak P, et al. Real-time hand gesture recognition using fine-  
tuned convolutional neural network[J]. Sensors, 2022, 22(3): 706.  
9. Tran D S, Ho N H, Yang H J, et al. Real-time hand gesture spotting and recognition using  
RGB-D camera and 3D convolutional neu- ral network[J]. Applied Sciences, 2020, 10(2):722.  
10. Mahmud H, Morshed M M, Hasan M K. A deep learning-based multimodal depth-aware dynamic  
hand gesture recognition[EB/OL]. arXiv:2107.02543, 2021.  
11. Ishihara T, Kitani  
K
M, Ma W C, et al. Recognizing hand-object interactions in wear- able  
camera videos[C]// IEEE International Conference on Image Processing. IEEE, 2015:  
1349–1353.  
12. Tekin B, Bogo F, Pollefeys M. Unified egocen- tric recognition of 3D hand-object poses[C]// IEEE  
Conference on Computer Vision and Pattern Recognition. IEEE, 2019: 4511–4520.  
13. Garcia-Hernando G, Yuan S, Baek S, et al. First-person hand action benchmark with RGB-  
D videos and 3D hand pose annota- tions[C]// IEEE Conference on Computer Vision and  
Pattern Recognition. IEEE, 2018: 409–419.  
14. Ahmad A, Migniot C, Dipanda A. Track- ing hands in interaction with objects: A  
review[C]// International Conference on Signal-Image Technology and Internet-Based  
Systems. IEEE, 2017: 360–369.  
15. Romero J, Kjellstr  
¨
o
m
H, Kragic D. Hands in action: Real-time 3D reconstruction of  
hands[C]// IEEE International Conference on Robotics and Automation. IEEE, 2010:  
458–463.  
16. Hamer H, Schindler K, Koller-Meier E, et al. Tracking a hand manipulating an object[C]//  
IEEE International Conference on Computer Vision. IEEE, 2009: 1475–1482.  
17. Kang B, Tan  
K
H, Jiang N, et al. Hand seg- mentation for hand-object interaction from depth  
map[C]// IEEE Global Conference on Signal and Information Processing. IEEE, 2017:  
259–263.  
18. Sridhar S, Mueller F, Zollh  
object[C]// European Conference on Com- puter Vision. Springer, 2016: 294–310.  
19. Cai M, Kitani M, Sato Y. Understanding hand-object manipulation with grasp types and  
¨ofer M, et al. Real- time joint tracking of a hand manipulating an  
K
object attributes[C]// Robotics: Science and Systems. 2016.  
20. Bertasius G, Park H S, Yu S X, et al. First person action-object detection with  
egonet[EB/OL]. arXiv:1603.04908, 2016.  
21. Schroder M, Ritter H. Hand-object interac- tion detection with fully convolutional net-  
works[C]// IEEE Conference on Computer Vision and Pattern Recognition Workshops.  
IEEE, 2017: 18–25.  
22. Yan W, Gao Y, Liu Q. Human-object inter- action recognition using multitask neural net-  
work[C]// International Symposium on Au- tonomous Systems. IEEE, 2019: 323–328.  
23. Kwon T, Tekin B, St  
u¨hmer J, et al. H2O: Two hands manipulating objects for inter- action  
recognition[C]// IEEE International Conference on Computer Vision. IEEE, 2021: 10138–  
Page 2493  
www.rsisinternational.org  
                                   
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
10148.  
pu  
24.  
K
¨
o
¨klu  
¨
O, Gunduz A, Kose N, et al. Real- time hand gesture detection and classifica-  
tion using convolutional neural networks[C]IEEE International Conference on Automatic Face  
and Gesture Recognition. IEEE, 2019: 1–8.  
25. Mujahid A, Awan M J, Yasin A, et al. Real- time hand gesture recognition based on deep learning  
YOLOv3 model[J]. Applied Sciences, 2021, 11(9): 4164.  
26. Lai K, Yanushkevich S N. CNN+RNN depth and skeleton based dynamic hand gesture  
recognition[C]// International Conference on Pattern Recognition. IEEE, 2018: 3451–3456.  
27. Pigou L, Van Den Oord A, Dieleman S, et al. Beyond temporal pooling: Recurrence and temporal  
convolutions for gesture recog- nition[J]. International Journal of Computer Vision, 2018,  
126(2): 430–439.  
28. Molchanov P, Gupta S, Kim K, et al. Hand gesture recognition with 3D convolutional  
neural networks[C]// IEEE Conference on Computer Vision and Pattern Recognition  
Workshops. IEEE, 2015: 1–7.  
29. Zhang L, Zhu G, Shen P, et al. Learning spa- tiotemporal features using 3DCNN and Con-  
vLSTM for gesture recognition[C]// IEEE International Conference on Computer Vi-  
sion Workshops. IEEE, 2017: 3120–3128.  
30. Gao Q, Chen Y, Ju Z, et al. Dynamic hand gesture recognition based on 3D hand pose  
estimation[J]. IEEE Sensors Journal, 2021, 22(18): 17421–17430.  
31. Miah A S M, Hasan M A M, Shin J. Dynamic hand gesture recognition using graph neural  
networks[J]. IEEE Access, 2023, 11: 4703–4716.  
32. Sun S. A survey of multi-view machine learn- ing[J]. Neural Computing and Applications,  
2013, 23(7): 2031–2038.  
¨
33. Shukla D, Erkent  
O
, Piater J. A multi-view hand gesture RGB-D dataset for human-  
robot interaction scenarios[C]// IEEE Inter- national Symposium on Robot and Human In-  
teractive Communication. IEEE, 2016: 1084–1091.  
34. Wang L, Ding Z, Tao Z, et al. Generative multi-view human action recognition[C]// IEEE  
International Conference on Computer Vision. IEEE, 2019: 6212–6221.  
35. Zhang Z, Wang C, Xiao B, et al. Cross-view action recognition using contextual maximum  
margin clustering[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2014,  
24(10): 1663–1668.  
36. Arnold E, Dianati M, De Temple R, et al. Cooperative perception for 3D object detec- tion  
in driving scenarios[J]. IEEE Transac- tions on Intelligent Transportation Systems, 2020,  
23(3): 1852–1864.  
37. Teepe T, Wolters P, Gilg J, et al. EarlyBird: Early fusion for multi-view tracking in bird’s-  
eye view[C]// IEEE/CVF Winter Conference on Applications of Computer Vision. IEEE,  
2024: 102–111.  
38. Gao Y, Maggs M. Feature-level fusion in per- sonal identification[C]// IEEE Computer So-  
ciety Conference on Computer Vision and Pattern Recognition. IEEE, 2005: 468–473.  
39. Fadadu S, Pandey S, Hegde D, et al. Multi-view fusion of sensor data for im- proved  
perception in autonomous driv- ing[C]// IEEE/CVF Winter Conference on Applications of  
Computer Vision. IEEE, 2022: 2349–2357.  
40. Seeland M, M  
¨ader P. Multi-view classification with convolutional neural networks[J]. PLoS  
One, 2021, 16(1): e0245230.  
41. Cheng J, Yin W, Wang K, et al. Adaptive fu- sion of single-view and multi-view depth for  
autonomous driving[C]// IEEE/CVF Con- ference on Computer Vision and Pattern  
Recognition. IEEE, 2024: 10138–10147.  
42. Zheng D, Zheng X, Yang L T, et al. Multi-view feature fusion network for cam- ouflaged  
object detection[C]// IEEE/CVF Winter Conference on Applications of Com- puter Vision.  
IEEE, 2023: 6232–6242.  
43. Ezati A, Dezyani M, Rana R, et al. A lightweight attention-based deep network via multi-  
scale feature fusion for multi- view facial expression recognition[EB/OL].  
arXiv:2403.14318, 2024.  
44. He K, Zhang X, Ren S, et al. Deep resid- ual learning for image recognition[C]// IEEE  
Page 2494  
www.rsisinternational.org  
                           
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026  
Conference on Computer Vision and Pattern Recognition. IEEE, 2016: 770–778.  
45. Lin T Y, Doll  
´
ar  
P, Girshick R, et al. Feature pyramid networks for object detection[C]//  
IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2017: 2117–2125.  
Page 2495  
www.rsisinternational.org