Agricultural Mechanized Picking and Sorting Technology Based on Object Detection and Improved MobileNet V1

Abstract

With the rapid development of computer vision technology, intelligent and mechanized agricultural production will inevitably introduce visual image technology to replace manual picking and sorting work. A two-stage detection and sorting integrated framework based on feature reuse is proposed to meet the dual requirements of high-precision target positioning and lightweight real-time classification for agricultural mechanized picking and sorting tasks. This framework uses the improved YOLOv8 as a spatial position perception module and enhances robustness to low-quality samples by introducing deep convolution and Wise IoU V3 loss function. At the same time, the adapted and improved MobileNet V1 is used as a lightweight mass sorting module, and Adam parameter optimization and Maxout activation function are adopted to improve classification efficiency. The results showed that in crop object detection in the experimental field, the mean detection accuracy of the YOLOV8+MobileNet V1 model robot was 0.954, and the average sorting error rate was 4.779%. The improved YOLOV8+MobileNet V1 model had an average sorting recognition accuracy of 0.983 and a picking rate of 0.933 in specific applications. The average picking time for each crop was 4.47 seconds, which was better than other models. The research findings denote that the intelligent crop picking and sorting method constructed by the research has high object detection and sorting accuracy, and the picking efficiency of the model is higher, which has good application prospects in mechanized agricultural production.