An Improved Waste Detection and Classification Model Based on YOLOV5

The improvement in people's lives has resulted in a significant rise in the amount of household garbage created on a daily basis, to the point where waste separation can no longer be disregarded, especially for the series of problems: manual waste classification is time-consuming and laborious, and human waste classification errors are caused by a lack of knowledge reserve related to waste classification. To address these issues, we propose a waste classification network YOLO-CG optimized on the basis of YOLOV5 network structure in campus scene. Firstly, YOLO-CG draws lessons from the optimization idea of Transformer performance improvement by stacking the ConvNeXt Blocks in the ratio of 3:3:9:3 as backbone, adding the big size kernel and other adjustments, upgrading the mean average precision (mAP) of the network model by 5%. Then, to maintain the original accuracy while reducing the number of parameters, a computationally reduced cheap operation is introduced, which employs a simple 3 * 3 convolution to achieve a low-cost acquisition of redundant feature maps, resulting in a reduction of 12% in parameter count while also increasing the mAP. Both theoretical analysis and experiments demonstrate the effectiveness of the improved network model.