Intelligent fruit and vegetable identification by the integration of lightweight YOLOv5 and SENet attention mechanisms

Abstract

Aiming at the problems of low efficiency of traditional manual picking in smart agriculture and insufficient recognition accuracy of existing algorithms in complex scenarios, an intelligent recognition method for fruit and vegetable information integrating lightweight YOLOv5 and SENet attention mechanisms is studied and proposed, and an intelligent recognition system for fruit and vegetable information is constructed. The system designs the myCPS module to replace the traditional CSP structure, adopts the MobileNet-v2 backbone network and combines mySwin Transformer to optimize feature extraction, and integrates the SENet mechanism to enhance the weights of key features. The results show that on the same test set, mAP reaches 99.78%, which is 14.55% higher than that of YOLOv5. The model size is 203.89 MB, which is reduced by 4.5% compared with the benchmark model YOLOv5. The floating-point operation amount is 3.32GFLOPs, which is 9.3% lower than that of the benchmark YOLOv5 model. The single-frame inference time of the model is 38.49 ms, which is 15.7% lower than that of YOLOv5. In practical applications, the CPU usage rate of the system is only 26%, and the processing time for the same task is 81.9% shorter than that of YOLOv5X. The system proposed in this research can effectively improve the accuracy and real-time performance of fruit and vegetable recognition, reduce labor costs, optimize the sorting and quality inspection processes of fruits and vegetables, and is of great significance for promoting agricultural automation and intelligence.