Improved lightweight peanut detection algorithm based on YOLO v3

Aiming at the problem of missed and wrong detection when the color sorting system detects granular crops, taking peanuts as a typical representative of granular crops, a lightweight peanut detection network based on YOLO v3 is proposed. The network mainly includes three-level controllable receptive field structure and feature map down-sampling structure. Among them, the three-level controllable receptive field structure is responsible for extracting the three types of receptive field information in the image, and the expansion of the receptive field can be controlled by adjusting the expansion rate of the hole convolution to achieve multi-scale detection; the feature map down-sampling structure includes a first-level downsampling layer and a second-level down-sampling layer. The first-level down-sampling layer is responsible for controlling the width change of the feature map of the middle layer, and can balance the contradiction between the accuracy of detection and the real-time nature. Tested on the constructed peanut detection data set, the improved lightweight peanut detection network based on YOLO v3 reached an average accuracy rate of 99.72%, and it took only 107ms to run on the Intel i5-9500 CPU. Compared with the test results of the YOLO v3 network, the average accuracy rate is increased by 5.51%, and the average running time is reduced by 96.3%. The experimental results show that the improved lightweight peanut detection network greatly improves the real-time performance of the YOLO v3 network on the CPU, while maintaining high detection accuracy.