Research on the Performance Improvement of YOLO Algorithm Based on C3 Module Optimization in Agricultural Harvesting

Abstract

The development of computer vision and deep learning has promoted agricultural automation. The YOLO series of algorithms are widely used in agricultural fields such as robot fruit picking, but still face challenges such as occlusion and light changes. This study is based on YOLOv5 6.1. The C3 module is lightweight processed based on the 5s model to obtain the C3-L module. In the experiment, the C3 module was replaced with C3-L at the positions of Backbone, Head and Backbone+Head respectively, and the CBAM and CA attention mechanisms were introduced to compare the performances of different models. The results show that the improved C3-L module can reduce resource invocation and graphics card memory usage during training. The stability of replacing the C3 module in the Head part is good. After adding the CBAM attention mechanism, the overall accuracy rate increases by 5%. When the accuracy rate requirement is not high, partially replacing the C3 module in the Backbone can reduce the call to hardware resources and decrease the video memory by 17.4%, which is conducive to operation in mobile hardware. This study provides a reference for the optimization of the YOLO algorithm in agricultural picking scenarios and its transplantation to devices such as microcontrollers.