Automated waste classification for smart recycling: A multi-class CNN approach with transfer learning and pre-trained models

This study presents a comparative analysis of pre-trained CNN models using transfer learning for the multi-class classification of recyclable waste. The dataset used is publicly available and comprises 15,150 images belonging to 12 different categories of recyclable household waste: paper, cardboard, biological, metal, plastic, green-glass, brown-glass, white-glass, clothes, shoes, batteries, and trash. The methodology involves implementing 14 different CNN models pre-trained on ImageNet (e.g., ConvNeXt_V1, ResNet50, MobileNet_V3_Large) on this dataset through a transfer learning procedure. In all pre-trained models, only the final classifier layer was trained, while all other layers were kept frozen, utilizing a feature extraction strategy that leverages previously learned visual representations. This approach significantly reduced training time and computational cost, while minimizing the risk of overfitting in scenarios with limited data. Model performance was evaluated using classification metrics such as accuracy, F1-score, precision, and recall, as well as computational efficiency indicators including architectural depth, parameter count, and training time. According to the findings, the ConvNeXt_V1 model achieved the highest performance with a test accuracy of 96.95 %. On the other hand, MobileNet_V2 (93.90 %), MobileNet_V3_Large (93.69 %), MNASNet_1_0 (93.69 %), RegNet_Y_800MF (92.87 %), and SqueezeNet_1_0 (92.10 %) offered shorter training times (ranging from 2856 to 7219 s) and stood out as efficient alternatives for embedded system applications. • Smart recycling via 12-class CNN waste recognition. • Frozen transfer learning reduces cost on CPU deployment. • ConvNeXt_V1 reaches 96.95 % accuracy and 95.72 % F1. • Lightweight CNNs enable fast embedded/mobile inference. • Class-wise error analysis exposes look-alike material confusions.