Establishment and Experimental Verification of Deep Learning Model for On-line Recognition of Field Cabbage

It is difficult to accurately identify and locate the target in the process of target application of field vegetables. Aiming at the problem, taking cabbage as the research object, the target recognition method and experimental research based on deep learning were carried out. Compared with Faster R-CNN, SSD and YOLO v5s, three current target detection networks with good performance, YOLO v5s was selected as the transfer learning model for field cabbage recognition. Lightweight neural network was selected as the backbone. The fusion depth can be separated and convoluted to reduce the calculation parameters of the model, YOLO-mdw network integrating MobileNet v3s backbone and deepwise separable convolution was proposed to realize the real-time recognition of field cabbage in complex environment. A cabbage target recognition and positioning method based on Kalman filter and Hungarian algorithm was proposed, and the model was deployed on NVIDIA Xavier NX development board. The experimental results showed that the recognition accuracy of the recognition network under sunny, cloudy and rainy weather conditions was 93.14%, 94.75% and 94.23%, respectively. The image processing time was 54.09 ms, which was 26.98% shorter than that of the original YOLO v5s. When the speed was not more than 0.6 m/s, the recognition accuracy was 94%, the average positioning error was 4.13 cm, and the average cabbage diameter recognition error was 1.42 cm. The designed target recognition system could identify and locate cabbage in complex field environment in real time, and provide technical support for target-oriented spraying. © 2022, Chinese Society of Agricultural Machinery. All right reserved.