Abstract: In view of the problem that the traditional online detection method of soybean combine crushing rate is time-consuming and labor-intensive by manual detection and is affected by subjective factors, an online detection method of the crushing rate of mechanically harvested soybean based on the DeepLabV3+ network was proposed. The online soybean image acquisition device was used to obtain the soybean image harvested by the combine harvester in real-time, and the images were labeled using the labeling software to construct the data set. To further improve the network training speed, the lightweight convolution network MobileNetV2 was selected to replace the Xception network in the backbone feature extraction network of the DeepLabV3+ network. In the prediction part, black edge cutting and splicing were used to improve the accuracy of image segmentation. The results showed that the comprehensive evaluation index F₁ of broken grain identification in the soybean sample image of the test set based on the DeepLabV3+ network model was 89.49%, and the comprehensive evaluation index F₁ of complete grain identification was 93.93%. The quantitative model of soybean crushing rate was established, and bench tests were carried out. The average relative error between the online detection method of soybean crushing rate and the average artificial detection result was 0.36%. This paper provides a reference for the online detection of the working quality of soybean combine harvester.