Abstract:Aiming at the problems of over-fitting and low accuracy caused by the accompanying cha-racteristics of weeds in cotton fields in Xinjiang, taking cotton seedlings and weeds in Xinjiang as research objects, the factors affecting the low recognition rate of weeds were analyzed, and a network recognition model based on Faster R-CNN was established. A total of 5 370 cotton seedlings and weeds were collected from different angles and different natural environments and different intensive levels. In order to ensure sample quality and diversity, color migration and data enhancement were used to improve the color characteristics of the image and to expand the sample size, and the network model training was performed in the PASCAL VOC format data set. By comprehensively comparing the recognition time and accuracy of the four networks VGG16, VGG19, ResNet50 and ResNet101, the VGG16 network training Faster R-CNN model was selected. On this basis, the optimal anchor scale with aspect ratio of 1∶1 was designed. Under this model, Xinjiang cotton seedlings and weeds were identified, and the average recognition accuracy of 91.49% was achieved. The average recognition time was 262 ms. It provides a reference for the development of agricultural intelligent precision weeding equipment.