Explainable Deep Transfer Learning for Automated Waste Classification: A Comparative Study Using Grad-CAM

Abstract

The study discussed the automated classification of waste with deep learning focusing on model interpretability. A comparison study of four image classifiers is performed. A custom CNN and three pretrained networks ResNet50, MobileNetV2, and EfficientNetB0 on the publicly available TrashNet* dataset containing 2527 images, and 6 waste categories. All models are preprocessed using data augmentation and evaluated using accuracy, precision, recall, and F1 score. The results indicate that transfer learning significantly improves the efficiency. EfficientNetB0 performed best among all with accuracy of approximately 84%, closely followed by ResNet50 (~83%). Whereas MobileNetV2 (~76%), while the custom CNN achieved nearly 51% accuracy. Our results are consistent with previous studies conducted on Trash-net showing performance improvement using transfer learning architectures such as ResNet and DenseNet models [7], [11], [14]. We have also used Gradient-weighted Class Activation Mapping (Grad-CAM) to explain the predictions of each model. The resulting heat maps show that the pre-trained models pay a lot of attention to useful regions of the waste object while the baseline CNN's attention is spread over a larger area, which is the reason for its being less successful. In the class-wise analysis, it is seen that the majority class paper is the easiest to classify, whereas the minority class trash is difficult, which highlights the issue of imbalance in the data sets. In summary, our work shows the effectiveness of transfer learning for waste recognition and the importance of explainable AI using Grad-CAM to interpret the behavior of deep models.