Page 2567

www.rsisinternaonal.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

Performance Analysis of Low-Computational Computer Vision

Techniques for Recyclable Waste Identification.

Awe Oluwayomi,

Dr. Enosegbe, Daniel Lucky,

Adeniyi Akanni

1,2,3

Department of Computer Science, Caleb University, Imota, Ikorodu, Lagos.

DOI:

https://doi.org/10.51583/IJLTEMAS.2026.150500206

Received: 08 May 2026; Accepted: 13 May 2026; Published: 15 June 2026

ABSTRACT

Waste management on the low-resource setting is also a serious issue because of the shortage of infrastructure,

and the lack of access to modern technologies, and the use of manual sorting of waste leads to environmental

pollution and ineffective recycling. Although deep learning-based image classification has demonstrated the

potential of automating waste, the majority of studies emphasize high-computation models, which cannot be

used in resource-constrained environments. The research gap is a significant lack of systematic comparisons on

both classification and computational efficiency of low-computational convolutional neural network (CNN)

models. This study aim to examine the appropriateness of lightweight CNN models in the identification of

recyclable wastes with a tradeoff between accuracy and resource consumption. Three low-computational CNN

models, namely EfficientNet-Lite, MobileNetV2, and SqueezeNet, were trained and tested using publicly

available datasets, such as TrashNet. Accuracy and F1-score were used to measure the classification

performance, and latency, memory usage, and model size were used to measure computational efficiency.

Findings show that EfficientNet-Lite had the best accuracy (92.4) and F1-score (92.0) but used more memory

and inference time whereas MobileNetV2 provided a reasonable compromise between performance and

efficiency. SqueezeNet was the lightest and the fastest model but with a lesser classification reliability. These

results give practical recommendations on the design of AI-powered waste management in low-resource settings.

This study helps to advance sustainable recycling efforts because it emphasizes the existence of the accuracy-

latency-memory trade-offs of lightweight CNNs and informs practitioners and researchers on the need to use the

right model whenever deploying lightweight CNNs to mobile, edge, or embedded applications in developing

regions.

Keywords: Recyclable Waste Identification, Low-Computational CNN, Computer Vision, Resource-

Constrained Environments.

INTRODUCTION

Proper waste separation and recycling is one of the major problems in the world especially with the increasing

rate of urbanization and population, which still increases the levels of solid waste (Ebikapade & Baird, 2016).

In a low-resource setting or any developing environmental setting, poor waste sorting practices are responsible

in environmental pollution, overuse of landfills, and loss of recyclable resources that could be used to build

circular economy projects. The manual waste segregation process is also prone to errors and labor intensive,

which can be easily scaled, thus demonstrating the necessity of automated and smart waste management

solutions that would help achieve more efficient and sustainable recycling. New technologies in computer vision

and artificial intelligence (AI) have already shown the high potential of it in waste recognition and sorting

(Nahiduzzaman et al., 2025). Image recognition systems with deep learning, particularly convolutional neural

networks (CNNs) have been shown to attain accuracy on a large scale in classifying between recyclable and non-

recyclable materials by learning intricate visual structures on large datasets. Such systems provide a bright future

of automated waste segregation where decision-making is quicker and human interactions are minimized in all

processes of waste management (Taglay et al., 2025). Yet, most of the state-of-the-art deep learning models are

computationally intensive, which demands large processing power, memory, and energy. These are limitations

that restrict their use in low-resource environments where high-end hardware, consistent power supply and

internet accessibility are frequently limited (Patil et al., 2024). Consequently, the implementation of the

Page 2568

www.rsisinternaonal.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

traditional deep learning models into the real-life waste management system in such settings is still not feasible.

This difficulty has encouraged the development of increasing interest in low-computational CNN methods to

trade-off classification performance and efficiency (Shahab et al., 2022). MobileNet, SqueezeNet and

EfficientNet-Lite are lightweight architectures that have smaller model size, low latency, and low memory

consumption, which is suitable to deploy on mobile and embedded devices. This study will attempt to undertake

a thorough performance study on low-computational computer vision methods of recyclable waste detection

(Bauravindah & Fudholi, 2024). The main contribution is the comparative analysis of the chosen lightweight

CNN models in terms of their accuracy, latency of inference, and memory footprint, thus giving useful

information on their applicability to resource-constrained waste management processes.

Problem Statement

The use of artificial intelligence in waste management is increasing, the adoption of high-performance AI

systems is still elusive in the developing and low-resource areas. The current waste classification solutions are

based on computationally expensive deep learning models and need such attributes as powerful hardware, large

memory capacity, stable electricity, and permanent access to the internet (Qu, 2022). These conditions drastically

reduce the application of these in actual waste segregation systems where resources are limited. As a result, such

environments still rely on manual sorting procedures that are not very efficient in waste identification, achieving

low results in recycling and causing even more harm to the environment (Sudha et al., 2016). The urgency is

thus to have waste classification models that are capable of sustaining reasonable accuracy whilst being run with

stringent computational requirements. One of the biggest challenges is the design of systems capable of operating

with a low amount of memory, low processing capabilities, and low power usage (Sayem et al., 2024). Unless

these limitations are tackled, the practical utility of AI-based waste recognition cannot be achieved in the areas

where it is most necessary. It fills this gap by considering the low-computational computer vision methods,

which provide an effective compromise between performance and resource efficiency in identifying recyclable

waste.

LITERATURE REVIEW

AI-Based Waste Classification Systems

The use of AI-based waste classification systems has been an emerging topic of research due to the increasing

investigations in automated methods of enhancing waste determine and recycling (Nasir & Al-Talib, 2023). The

majority of existing works use image-based methods that rely on computer vision and deep learning algorithms

to detect and classify various types of waste materials which include plastic, paper, metal, glass, and organics

(Malik et al., 2022). The leading models in this field are convolutional neural networks (CNNs), which are

characterised by high levels of discriminative visual features of waste images. Simple architectures like

VGGNet, ResNet, Inception, and DenseNet have been extensively used and they have been found to have high

classification accuracy when trained on standard datasets like TrashNet, TACO and WasteNet (Mao et al., 2021).

Several research have examined how transfer learning can be used, in which pre-trained CNNs can be fine-tuned

to waste data to minimize training time and enhance performance. Such methods have shown promising

performances especially in restricted environments where there exists adequate computation resources (Kartik

et al., 2023). Moreover, data augmentation methods are usually employed to enhance the model resistance to

changes in the waste appearance and include rotation, scaling, and lighting reconfiguration. Studies that are even

more recent have started to deal with the aspect of practical deployment with the introduction of lightweight and

mobile-friendly waste classification models (William et al., 2024). MobileNet, SqueezeNet and EfficientNet

variants of architectures have been explored due to their smaller parameter size and accelerated inference speed.

It has also been observed that some of the studies have incorporated waste classification models into smart bins,

mobile applications as well as edge devices in order to support real-time waste sorting (Frost et al., 2025

). Nonetheless, in spite of these improvements the current literature tends to focus more on the accuracy of the

classification rather than the computation efficiency. Deep comparative studies that collectively assess accuracy,

latency, and memory consumption are still scarce, especially regarding low-resource settings. Such a gap

demonstrates the necessity of systematic performance appraisal of the low-computational AI-based waste

classification systems.

Page 2569

www.rsisinternaonal.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

Review of Related Works

In this section, this study examine some of the previous approaches used by researchers for Image Classification

of Recyclable Waste for Low-Resource Environments. Below are brief review of research studies that have been

conducted using these approaches.

Yadav (2025) proposes an AI-driven system to improve urban waste management through automated waste

recognition, sorting, and optimized collection routing. The system applies deep learning techniques, particularly

convolutional neural networks, to classify recyclable materials from image data with high accuracy. Designed

for low-power devices, it is suitable for resource-constrained environments. Image augmentation and

normalization enhance model reliability. Results show improved sorting efficiency, reduced landfill use,

optimized collection schedules, and lower environmental impact. The study demonstrates that AI can modernize

urban recycling systems and supports its adoption for sustainable waste management and resource conservation.

Sivakumar et al. (2025) propose a CNN-based automated waste classification system to improve recycling

efficiency and reduce reliance on manual sorting. The model is trained on diverse waste images to enhance

generalization and uses data augmentation to prevent overfitting. Results show high classification accuracy,

reduced sorting errors, and improved waste management efficiency. The study concludes that CNN-based

artificial intelligence systems are reliable tools for waste classification and can significantly enhance recycling

processes while contributing to environmental protection and sustainable waste management.

Hossen et al. (2024) propose RWC-Net, a deep learning model for classifying six waste categories using the

TrashNet dataset. Trained with data augmentation and optimized using Adam, the model achieved 95.01%

accuracy, outperforming established architectures such as ResNet50 and MobileNet-v2. High F1-scores across

all classes demonstrate robustness. Score-CAM enhances interpretability by highlighting decision regions. The

study confirms RWC-Net’s reliability for automated waste sorting and recommends extending it to more waste

types and real-world recycling applications.

Ahmad, Khan, and Al-Fuqaha (2020) propose a double-fusion approach to improve image-based waste

classification by combining features and outputs from multiple CNNs, including AlexNet, VGGNet, GoogleNet,

and ResNet. Using the TrashNet dataset, the method integrates early and late fusion strategies with optimized

weighting through particle swarm optimization. Results show a 3.58% accuracy improvement over single models

and conventional fusion techniques. The study demonstrates that multi-CNN feature integration significantly

enhances classification reliability for automated waste management systems.

Sayem et al. (2024) propose a deep learning-based system to automate waste sorting and improve recycling

efficiency. Using CNNs trained on diverse waste images with augmentation and object detection techniques, the

system accurately classifies and detects various waste types. Evaluated via accuracy, precision, recall, and F1-

score, it demonstrates high reliability and performance. The approach reduces human labor, enhances sorting

efficiency, and can be applied in automated recycling facilities, showing that deep learning effectively supports

sustainable and efficient waste management practices.

Sallang et al. (2021) develop a CNN-based waste classification system using TensorFlow Lite for lightweight

deployment on IoT devices with LoRa-GPS for real-time location monitoring. Trained on diverse waste images,

the model accurately classifies waste while enabling remote tracking and management. Designed for smart cities,

it reduces resource usage and automates sorting. The combination of lightweight deep learning and IoT improves

waste collection efficiency, provides immediate data for monitoring, and supports scalable, low-power smart

waste management systems with potential for further expansion.

Chhabra et al. (2024) develop a DCNN-based system to classify household waste into organic and recyclable

materials, improving on standard CNNs through hyperparameter tuning and architectural enhancements. Using

25,077 images, the model achieved 93.28% accuracy with low MDR (2.6%) and FDR (4.5%), outperforming

VGG16, VGG19, MobileNetV2, DenseNet121, and EfficientNetB0. The system automates sorting, reduces

manual labor, promotes sustainability, and enhances recycling efficiency, demonstrating that advanced deep

learning models can effectively support smart waste management in urban environments.

Page 2570

www.rsisinternaonal.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

Malik et al. (2022) develop an EfficientNet-B0-based model to classify municipal solid waste into bio, plastic,

glass, metal, and paper. Optimized with transfer learning and trained on ImageNet, the system achieves accuracy

comparable to EfficientNet-B3 but with fewer parameters and greater efficiency. Lightweight and resource-

efficient, it can operate in varied environments, enabling automated, adaptable waste sorting. The model

enhances waste management, supports deployment in resource-limited settings, and shows potential for further

improvement to handle more waste types and real-time applications.

Yang and Li (2020) develop WasNet, a lightweight CNN for accurate waste classification in low-resource

environments. Integrated with smart bins, a mobile app, and a centralized platform, it automates waste sorting

and management. Using data augmentation and attention modules (SE, CBAM), WasNet achieves high accuracy

on multiple datasets (TrashNet 96.1%, Huawei 82.5%). With only 1.5 million parameters, it outperforms

ShuffleNet-V2 and MobileNetV3-Small. The system enhances sustainable waste disposal, reduces manual labor,

and enables efficient municipal waste monitoring and decision-making.

METHODOLOGY

Dataset Description

To assess the performance of computer vision-based approaches with low computational requirements for the

identification of the recyclable waste in low resources environments, this study used publicly available

recyclable waste image datasets, which is mostly the TrashNet dataset. The TrashNet dataset consists of labeled

waste images of types plastic, paper, cardboard, glass, metal, and organic waste materials, divided into two

classes: recyclable and non-recyclable. The data set was chosen due to its visual diversity, multiple type of waste,

and balanced representation, which is ideal for supervised image classification tasks. The data set was

preprocessed with quality assessment to exclude the blurred, duplicate, low-resolution, corrupted image data that

may have a negative impact on the model learning process before experimentation. The final dataset distribution

was designed to be representative in terms of waste category and minimize classification bias to dominant waste

classes. About 70% of the pictures were used for training the model, 15% for validation, and 15% for testing.

That dataset splitting was stratified was carried out in order to ensure that the proportions of each waste category

were the same across each of the subsets. The training dataset was used to learn the features and optimize the

parameters, the validation dataset was used to tune hyper-parameters and avoid overfitting when training. The

testing data set was set aside specifically to assess the performance of the final model as an unbiased testing of

the model. Also, some other augmentation techniques were added to enhance the diversity of the dataset and the

generalization of the model, particularly for the less numerous categories of waste. This pre-processing stage

enabled the models to adequately classify recycled materials in different environmental conditions which can be

encountered in waste management systems in practice.

Image Preprocessing Methods

To enhance the quality of images, standardise the input size, minimise computational complexity and increase

the robustness of the lightweight CNN models, the images were processed. The datasets consist of images taken

with varying lighting, background, orientation and resolution so there was a need to preprocess the data to make

it more uniform for model training. To bring all images to a common size, the first pre-processing step was to

resize them all to 224 × 224 pixels required by MobileNetV2, SqueezeNet and EfficientNet-Lite architectures.

Resize also shortened training and inference time, and memory usage. After resizing, the intensity values of the

pixels were normalized by scaling the values between 0 and 1. This normalization has enhanced the numerical

stability in the optimization process and faster CNN model convergence. Images that were visually distorted or

of low quality, which might cause classification mistakes, were filtered out. Images of waste were selectively

processed with gaussian smoothing and image sharpening to highlight the objects and edges in the waste images.

In order to enhance the generalization of the model and reduce overfitting, several data augmentation techniques

were used on the images to be used for training. These were random rotation, horizontal flipping, zoom scaling,

brightness adjustment, cropping and translation transformations. To simulate object positioning variations,

rotations of −20° to +20° were applied to the objects; brightness adjustments allowed models to adapt to different

lighting conditions that typically occur outdoors during waste disposal. The data set was further varied by

flipping the images horizontally and performing zoom transformations, which further enhanced the ability of the

Page 2571

www.rsisinternaonal.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

model to identify recyclable materials from various angles.Data was further varied by horizontally flipping the

images and performing Zoom transformations, which further enhanced the ability of the model to identify the

recyclable materials from various angles. Only the training set was data augmented to avoid information leakage

during validation and testing. All CNN models were given consistent and optimised image inputs using the pre-

processing pipeline, which can be implemented in a real-time resource-constrained system for lightweight

classification.

Model Selection and Training Configurations

The study focused on three lightweight convolutional neural network architectures, namely MobileNetV2,

SqueezeNet, and EfficientNet-Lite, because of their low computational requirements, reduced parameter sizes,

and suitability for deployment on mobile, embedded, and edge devices. MobileNetV2 was chosen since it

incorporates the depthwise separable convolutions and inverted residual blocks, which help to control the

number of computational operations without sacrificing the classification accuracy. SqueezeNet was chosen

because of its very small parameters with its fire modules that can cut down on the complexity of the network

with no significant loss of performance. Its efficient network depth, width and resolution for efficient feature

extraction allowed EfficientNet-Lite to be selected. Experiments were performed in Python and TensorFlow Lite

as the main frameworks to implement the model. The reason for using TensorFlow Lite is its ability for

lightweight deployment on low-resource mobile and embedded systems. Supporting libraries employed for

image processing, image visualization and evaluation procedures include NumPy, OpenCV, Matplotlib and

Scikit-learn. To simulate the real-life deployment scenario of such edge devices, the training experiments were

carried out on a middle-of-the-road laptop with an Intel Core i7 processor and 16 GB of RAM. The processing

parameters and the way the datasets were split were the same for both CNN models to ensure an apples-to-apples

comparison. The lightweight CNN models were used with transfer learning, where the models pretrained on the

ImageNet dataset were fine-tuned on the recyclable waste dataset. This method not only saved training time, but

also enhanced classification accuracy, particularly when the number of data points was small. For training, the

final classification layers were replaced by customized fully connected layers, based on waste categories of the

dataset. Adam optimizer has been employed due to its adaptive learning feature and its consistent convergence.

The starting learning rate was 0.001 and the number of images per batch was 32. To avoid overfitting and

enhance the stability of models, early stopping and the learning rate reduction callbacks were applied. The

training was performed for 50 epochs but was automatically stopped after a few epochs if the validation loss did

not show any improvement. The fully connected layers also used the dropout method to prevent overfitting and

to enhance the generalization of the model. To determine the best configuration for each lightweight CNN model,

hyperparameter tuning was performed, and the validation loss and the F1-score were used as optimization

function.

Experimental Setup and Deployment Simulation

The experimental setup was designed to simulate realistic deployment scenarios for recyclable waste

identification systems deployed in low-resource environments. The light-weight CNN models were assessed not

only with respect to the classification accuracy, but also with respect to the computational efficiency under a

hardware limit, which is mostly observed in the developing region. The experiments were carried out in two

computational settings. The first environment was a standard laptop system with which to train the model and

perform initial testing, and the second environment simulated edge deployment with light computational

constraints. During deployment simulations, TensorFlow Lite optimization techniques like model quantization

and reduced precision inference were used to reduce memory usage and inference latency. The processing of

image streams sequentially in real time was used to perform real-time inference experiments, thus emulating

automated waste sorting operations in smart bins and recycling facilities. The system performed a classification

speed, memory allocation, processor utilization and model loading times measurement during the inference

process. These experiments were conducted to assess the feasibility of the light weight CNN models in the real

world waste management system having less hardware.

Page 2572

www.rsisinternaonal.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

Evaluation Procedures and Performance Metrics

The evaluation process aimed to evaluate the classification performance and the computational efficiency of the

lightweight CNN models for the purpose of waste identification in recyclable waste streams. To make the model

performance comparison comprehensive, a mix of the classification metrics and resource-efficiency metrics

were used. The metrics used to assess the classification performance were accuracy, precision, recall and F1-

score. The accuracy was calculated as the total percentage of correctly classified waste images, and precision as

the number of correctly identified recyclable images divided by the number of all predicted recyclable output

images. The models were evaluated using recall to correctly identify the actual instance of recyclable waste and

precision and recall were measured using the F1-score. To visualize the classification results for all the waste

components, the confusion matrix was created for each CNN model. Also, to identify the pattern of

misclassification between the recyclable waste component, plastic, paper, and glass, the confusion matrix was

created. Latency, memory usage and model size were measured to assess the computational efficiency. Latency

was the average time (in milliseconds) it took to classify an image and was measured in milliseconds. The amount

of RAM used during the inference process was used as a measure of memory usage, and the overall memory

footprint of each lightweight CNN architecture was used as a measure of model size. These statistics were

essential because in low resource areas, often there are computational and storage constraints. Multiple

experimental runs were performed under the same conditions and average values reported to enhance the

repeatability and reliability of the results. The experiments were performed on the same hardware and processing

pipeline with the same dataset partitionings to fairly compare the models across the architectures. The evaluation

framework hence offered a clear and repeatable method of evaluating the applicability of low-computational

computer vision methods to the identification of recyclable waste in resource-poor settings.

RESULTS

Table 1: Classification Performance Comparison

CNN Model

Accuracy (%)

Precision (%)

Recall (%)

F1-Score (%)

MobileNetV2

91.2

90.5

91.0

90.7

SqueezeNet

87.5

86.8

87.0

86.9

EfficientNet-Lite

92.4

92.0

92.1

92.0

Figure 1: graph of Classification Performance Comparison

Figure 1 shows the performance of three low-computational CNN models, which are MobileNetV2, SqueezeNet,

and EfficientNet-Lite, in classifying recyclable wastes. EfficientNet-Lite attains the best performance with

92.4% accuracy and 92.0% F1-score and, thus, a better balance between the precision and the recall.

MobileNetV2 has a slightly lower accuracy of 91.2 and F1-score of 90.7, showing a competitive capability of

classification. SqueezeNet has the lowest values, 87.5% accuracy and 86.9% F1-score, but it has the advantage

of lightweight architecture. In general, the graph shows that there is a trade-off point between the complexity of

Page 2573

www.rsisinternaonal.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

models and performance, with EfficientNet-Lite offering the best classification performance when deployed in

low-resource settings.

Table 2: Evaluation Metric

CNN Model

Accuracy

(%)

F1-Score

(%)

Latency (ms)

Memory

Usage (MB)

Model Size

(MB)

MobileNetV2

91.2

90.7

120

SqueezeNet

87.5

86.9

EfficientNet-

Lite

92.4

92.0

140

120

Table 2 shows trade-offs between classification and computational efficiency between low-computational CNN

models used to identify recyclable waste. EfficientNet-Lite has the highest accuracy (92.4%), F1-score (92.0%),

and is the most efficient in terms of classification, 22 ms latency, 140 MB memory use, and 20 MB model size.

SqueezeNet is the smallest, and its latency (12 ms), memory (90 MB), and model size (5 MB) are the lowest,

but the accuracy (87.5) and F1-score (86.9) are worse. MobileNetV2 offers a trade-off between performance

(91.2% accuracy) and moderate latency and memory usage, which enables it to be deployed successfully in

limited resources.

DISCUSSION

The findings show evident differences in performance of the low-computational CNN models that have been

evaluated. EfficientNet-Lite had the best accuracy (92.4), F1-score (92.0), which implies that it has a high

capacity of detecting recyclable waste with the least misclassifications. The difference in its performance and

computational requirements (slightly higher latency of 22 ms and memory consumption of 140 MB) indicates

the trade-off but implies that this model may be deployed in the environment with moderate hardware

requirements. MobileNetV2, with 91.2 and 90.7 accuracy and F1-score, provides a compromise between the

performance of classification and efficiency as well as it is very appropriate in a mobile environment or edge

deployments where resources are more likely to be a bottleneck. Although SqueezeNet is the fastest and lightest

model (12 ms latency, 90 MB memory, 5 MB model size), it is the least accurate (87.5%) and F1-score (86.9%),

meaning that it might not work well in settings where there is a need to ensure high classification reliability.

These results are in line with existing literature that lightweight CNN models can achieve acceptable accuracy,

much smaller model size and lower computational cost. The past studies tended to emphasise one of the two;

accuracy or efficiency, though this paper emphasises the need to consider both together especially in cases of

low resources deployment. In a more pragmatic viewpoint, these models allow real-time or near real-time

recyclable wastes to be identified in an environment with inferential computing capability, like developing areas

or intelligent recycling receptacles. EfficientNet-Lite can be used with centralized or semi-automated sorting

systems, MobileNetV2 can be implemented in handheld devices or mobile devices, and SqueezeNet can be

deployed on devices with limited resources when speed and memory consumption are extremely important. In

general, these findings can be used to implement practical suggestions to create sustainable waste management

systems based on AI.

CONCLUSION

This study compared the results of low-computational convolutional neural network (CNN) models, namely,

EfficientNet-Lite, MobileNetV2 and SqueezeNet in recyclable waste recognition in low-resource settings. These

findings suggest that EfficientNet-Lite has the best classification accuracy (92.4%), F1-score (92.0%), which

proves its high potential to identify correctly the recyclable materials. MobileNetV2 provided a trade-off on

performance and efficiency with a slightly lower accuracy (91.2) but medium latency and memory consumption

due to which it could be deployed to a mobile or edge environment. Although SqueezeNet has the highest

inference (12 ms) and minimum memory footprint (90 MB), its accuracy (87.5%) was relatively lower, which

is an indicator of the trade-off between the simplicity of the model and its ability to classify. This study adds to

the AI and sustainable waste management by delivering a comparative analysis of the low-computational CNNs

Page 2574

www.rsisinternaonal.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

on a systematic basis, focusing on the classification and computational efficiency. These lessons are especially

useful when developing AI-based waste sorting systems in resource-limited environments, i.e. developing

countries or smart bins. This study helps to show that scalable, inexpensive, and sustainable solutions to

automated recycling can be achieved by showing how low-weight models can achieve high accuracy and reduce

hardware needs. To practitioners, the results suggest the use of CNN models depending on deployment factors:

EfficientNet-Lite with semi-automated systems or centralized systems, MobileNetV2 with mobile apps, and

SqueezeNet with ultra-low-resource systems. In the eyes of researchers, future studies may include hybrid or

quantized models, larger and more diverse data and combination with multimodal sensors to further improve the

performance, reliability, and applicability in real world waste management situations.

REFERENCE

1. Ahmad, K., Khan, K., & Al-Fuqaha, A. (2020). Intelligent fusion of deep features for improved waste

classification. IEEE Access, 8, 96495-96504. Retrieved from

https://doi.org/10.1109/ACCESS.2020.3023092.

2. Bauravindah, A., & Fudholi, D. (2024). Lightweight steganalysis models: MobileNet, ShuffleNet,

EfficientNet. Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi), 8(6), 737-747.

https://doi.org/10.29207/resti.v8i6.6091

3. Chhabra, M., Sharan, B., Elbarachi, M., & Kumar, M. (2024). Intelligent Waste Classification Approach

Based on Improved Multi-Layered Convolutional Neural Network. Multimedia Tools and Applications,

83, 84095–84120. https://doi.org/10.1007/s11042-024-18939-w

4. Ebikapade, A., & Baird, J. (2016). Understanding waste and waste management. Journal of Management

and Sustainability, 6(4), 88. https://doi.org/10.5539/jms.v6n4p88

5. Frost, A., Jean, G., & Ailyn, D. (2025). Lightweight CNNs for edge computing in resource-limited

settings.

Retrieved from https://www.researchgate.net/publication/395767938

6. Hossain, M. A., Ferdous, N., & Ferdous, E. (2024). Crisis-driven Disruptions in Global Waste

Management: Impacts, Challenges, and Policy Responses Amid COVID-19, Russia-Ukraine War,

Climate Change, and Colossal Food Waste. Environmental Challenges, 14, 100807. Retrieved from

https://doi.org/10.1016/j.envc.2023.100807.

7. Kartik, K., Ahmed, T., Ghosh, S., Gupta, R., & Tripathi, A. (2023). Transfer learning with CNNs on small

datasets. In Proceedings of the International Conference on Machine Learning and Applications (pp.

1087-1099). CC BY 4.0 license.

8. Malik, M., Sharma, S., Uddin, M., Chen, C. L., Wu, C. M., Soni, P., & Chaudhary, S. (2022). Waste

Classification for Sustainable Development Using Image Recognition with Deep Learning Neural

Network Models. Sustainability, 14(12), 7222. Retrieved from https://doi.org/10.3390/su14127222.

9. Mao, W., Chen, W., Wang, C., & Lin, Y. (2021). Optimized CNN for recycling waste classification.

Resources, Conservation and Recycling, 164, 105132. https://doi.org/10.1016/j.resconrec.2020.105132

10. Nahiduzzaman, M., Ahamed, M., Naznine, M., Karim, M., Kibria, H., Ayari, M., Khandakar, A., Ashraf,

A., Ahsan, M., & Haider, J. (2025). Deep learning for automated waste classification. Knowledge-Based

Systems, 310, 113028. https://doi.org/10.1016/j.knosys.2025.113028

11. Nasir, I., & Al-Talib, G. (2023). AI techniques for waste classification: A literature review. Technium:

Romanian Journal of Applied Sciences and Technology, 5, 49-59.

https://doi.org/10.47577/technium.v5i.8345

12. Patil, D., Rane, N., Desai, P., & Rane, J. (2024). Machine learning and deep learning: An overview. In

Machine Learning and Deep Learning (pp. 11-30). ISBN: 978-81-981367-4-9.

13. Qu, D. (2022). AI in university waste classification management. In Advances in Artificial Intelligence

and Applied Mathematics (pp. 287-296). Springer, Cham. https://doi.org/10.1007/978-3-030-97196-

0_27

14. Sallang, N. C. A., Islam, M. T., Islam, M. S., & Arshad, H. (2021). A CNN-based smart waste

management system using TensorFlow lite and LoRa-GPS shield in Internet of Things environment.

IEEE Access, 9, 153560-153574. Retrieved from

https://doi.org/10.1109/ACCESS.2021.3055128.

15. Sayem, F. R., Islam, M. S. B., Naznine, M., Nashbat, M., Hasan-Zia, M., Kunju, A. K. A., ... &

Chowdhury, M. E. (2024). Enhancing Waste Sorting and Recycling Efficiency: Robust Deep Learning-

Page 2575

www.rsisinternaonal.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

Based Approach for Classification and Detection. Neural Computing and Applications, 1–17. Retrieved

from https://doi.org/10.1007/s00521-024-06994-8.

16. Shahab, S., Anjum, M., & Umar, M. (2022). Deep learning in solid waste management: A review.

International Journal of Advanced Computer Science and Applications, 13.

https://doi.org/10.14569/IJACSA.2022.0130347

17. Sudha, S., Vidhyalakshmi, M., Pavithra, K., Sangeetha, K., & Swaathi, V. (2016). Deep learning for eco-

friendly waste classification. In 2016 International Conference on Technological Innovations in

Management, Engineering, and Sciences (TIARE) (pp. 65-70). IEEE.

https://doi.org/10.1109/TIAR.2016.7801215

18. Taglay, D., Rigor, J., Royo, R., & Fernandez, R. (2025). CNN-based image recognition for recyclables.

International Journal of Latest Technology in Engineering Management & Applied Science, 14, 606-617.

https://doi.org/10.51583/IJLTEMAS.2025.1409000073

19. William, Lily & Ayodele, Harry & Blake, Harrison. (2024). Evaluating the Impact of Data Augmentation

Techniques on Model Performance.

https://www.researchgate.net/publication/389357090_Evaluating_the_Impact_of_Data_Augmentation_

Techniques_on_Model_Performance/citation/download

20. Yang, Z., & Li, D. (2020). WasNet: A neural network-based garbage collection management system.

IEEE Access, 8, 103984-103993. Retrieved from https://doi.org/10.1109/ACCESS.2020.3021510.