Deep learning has emerged as an effective tool for automatic waste classification, supporting cleaner cities and more sustainable recycling systems. Because environmental protection is central to the United Nations Sustainable Development Goals (SDGs), improving the sorting and processing of everyday waste is a practical step toward this broader objective. In many real-world settings, however, waste is still sorted manually, which is slow, labor-intensive, and prone to human error. Although convolutional neural networks (CNNs) can automate this task with high accuracy, many state-of-the-art models remain too large and computationally demanding for low-cost edge devices intended for deployment in homes, schools, and small recycling facilities. In this work, we investigate lightweight waste-classification models suitable for TinyML deployment while preserving competitive accuracy. We first benchmark multiple CNN architectures to establish a strong baseline, then apply complementary compression strategies including quantization, pruning, singular value decomposition (SVD) low-rank approximation, and knowledge distillation. In addition, we evaluate an RL-guided multi-teacher selection benchmark that adaptively chooses one teacher per minibatch during distillation to improve student training stability, achieving up to 85% accuracy with only 0.496 M parameters (FP32 ≈ 1.89 MB; INT8 ≈ 0.47 MB). Across all experiments, the best accuracy–size trade-off is obtained by combining knowledge distillation with post-training quantization, reducing the model footprint from approximately 16 MB to 281 KB while maintaining 82% accuracy. The resulting model is feasible for deployment on mobile applications and resource-constrained embedded devices based on model size and TensorFlow Lite Micro compatibility.

Insect bites are a common cause of skin irritation and can contribute to disease transmission through vector-borne pathogens. Early identification of the likely biting organism can assist preliminary guidance (e.g., monitoring for warning signs, considering exposure history) and may reduce complications through timely follow-up. This paper studies a compact attention-guided learning framework for multiclass insect-bite image classification under strict storage constraints. A teacher network (BiteAI-T) based on MobileNetV3-Small is trained with spatial attention pooling to emphasize lesion-relevant regions while maintaining an efficient backbone. A lightweight depthwise-separable student (BiteAI-S) is trained using multi-level knowledge distillation that combines softened-logit matching with intermediate supervision through attention-map alignment and pooled-feature matching. Model storage is further reduced through weight-only quantization-aware training using an LSQ-inspired learnable scaling factor; BatchNorm running statistics are frozen during quantization fine-tuning to improve stability. Experiments on an eight-class dataset (ants, bed bugs, chiggers, fleas, mosquitos, no bites, spiders, ticks) show that BiteAI-T reaches 93.75% test accuracy. For deployment, we export (i) a TorchScript Lite teacher artifact (BiteAI-T_Lite, 2.35 MB) and (ii) a weight-only int8 student artifact (BiteAI-S_int8, 0.992 MB). Comparative results are also reported for an SVD-compressed + fine-tuned FP16 variant (92.66% test accuracy, 2.84 MB), illustrating accuracy–size trade-offs across compression strategies.

Automatic waste classification is an important enabling technology for cleaner cities, source-level recycling, and low-cost smart-bin systems. Although modern convolutional neural networks achieve strong recognition performance, their deployment on affordable edge devices remains constrained by memory footprint, computational cost, and response latency. This paper presents an edge-oriented compact CNN framework for waste image classification, combining a high-accuracy MobileNetV4 reference model with three lightweight student architectures: EfficientNet-Lite0, LCNet-0.5, and MobileNetV3-Small-0.5. All models are evaluated on TrashNet under a unified preprocessing, training, and size-accounting protocol, allowing a clear comparison of accuracy–efficiency trade-offs. On the main stratified train/validation/test split, the MobileNetV4 teacher achieves 97.09% top-1 accuracy, while the compact students retain strong performance with substantially smaller footprints: EfficientNet-Lite0 reaches 93.99% with 3.38 M parameters, LCNet-0.5 reaches 94.18% with only 0.61 M parameters, and MobileNetV3-Small-0.5 reaches 87.73% with 0.57 M parameters. A complementary stratified five-fold evaluation, including both knowledge-distilled and non-distilled student variants, provides a robust assessment of model behavior across data partitions and confirms LCNet-0.5 as the most suitable sub-megabyte candidate under the proposed size–accuracy selection rule. The selected LCNet-0.5 model achieves a macro-F1 score of 0.9247 on the main TrashNet test split and is integrated into a self-contained Raspberry Pi 3 Model B+ prototype that performs local camera-to-display inference with an observed end-to-end latency of approximately 1.0 s per image. Cross-dataset evaluation on RealWaste further shows that the compact model can be adapted effectively to cluttered real-world imagery through short fine-tuning. Overall, the results demonstrate that careful lightweight architecture selection, supported by knowledge distillation analysis and edge-prototype validation, can deliver accurate, compact, and practically deployable waste classifiers for resource-constrained environments.