Attention-based Improved YOLOv5 for Effective Waste Detection and Classification System

Deep learning architectures are now prevalent among several application domains due to their better performance. Waste detection and classification are two applications that can be enhanced using deep learning. This work focuses on vision-based analysis using only camera sensors; other sensors are not included due to cost optimization. Different waste detection approaches were earlier designed based on computer vision and image processing, but there is a scope for improvement in detection accuracy. With this aim, this work proposes an improved YOLOv5 architecture. This architecture makes use of both spatial and channel attention to enhance the performance of the suggested model and extract more pertinent information during feature learning. The waste recycling plant dataset (WaRP-D) is used to analyze the performance of the improved architecture. This dataset contains image data of different waste objects such as bottles, containers, cans, cardboard, etc.; their annotations are also publicly available. The performance of our proposed model is measured considering various parameters in terms of precision, recall, and mean average precision with different thresholds (mAP<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0.5</sup> and mAP<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0.5-0.95</sup>). A comparison of the proposed architecture is carried out using simple YOLOv5 architecture, which shows the effectiveness of the proposed approach. This work also compares the classification performance of this improved YOLOv5 architecture with gated recurrent units (GRU) and an optimized GRU using grey wolf optimization (GWO-GRU) on the Waste Segregation Image dataset.