Plastic Waste Classification Based on KEffNet-B1 V2 with Dual Attention Integration

This research presents an innovative approach for plastic waste classification using the kEffNetV2-B1 model enhanced with Dual Attention Mechanism. The dataset utilized comprises 9,000 images of plastic waste, segmented into 7,200 images for training and 1,800 images for validation. The kEffNetV2-B1 model employs transfer learning techniques, leveraging pre-trained weights from the ImageNet dataset, while allowing fine-tuning of the last ten layers to adapt the model for our specific classification task. The model was evaluated with various configurations, employing 2, 4, 8, 16, and 32 channels per group, demonstrating an accuracy range of 98.78% to 99.28% across different configurations with Dual Attention. In contrast, the baseline model, EfficientNetV2-B1, achieved an accuracy of 96.61%. Moreover, when tested without the Dual Attention Mechanism, the model still achieved high accuracy, ranging from 98.61% to 98.83%. The results underscore the effectiveness of integrating attention mechanisms and transfer learning techniques in enhancing model performance for image classification tasks in the context of environmental sustainability. Future work will focus on further optimizing the model and exploring additional data augmentation techniques to improve classification accuracy.