Abstract: Abstract: Soybean is one of the most important oil crops in the national grain and oil security system. Mechanized harvesting can be the top priority to improve the level of soybean production. Among them, the crushing rate and impurity rate of soybean seeds are the important performance indexes to evaluate the soybean combine harvesters. The real-time data of crushing rate and impurity rate can greatly contribute to realizing the intelligent control of soybean combine harvesters. However, the current manual detection of soybean cannot fully meet the requirement in the process of combined harvester operation. Particularly, the manual operation was only made after shutdown, due to the high misjudgment rate and low efficiency. This study aims to realize the online detection of soybean grain for the crushing rate and impurity rate during mechanical harvesting of soybean using an improved U-net network. Taking the real-time soybean image harvested by a soybean combine harvester as the object, the open source annotation software Labemel was used to annotate and construct the basic data set. The U-net network structure was combined with the VGG16 network, Batch Normalization (BN) before each Rectified Linear Unit (ReLu). The overfitting was avoided, due to the soybean image adhesion, stacking and complex semantic information. The convolution block attention module (CBAM) was added to the feature map extracted from the encoder, in order to suppress the activation of the irrelevant region for the less redundant part. The up-sampling of the nearest neighbor interpolation was used to replace the decoder with the transpose convolution, where the checkerboard effect was caused by the transpose convolution. A comparative test was carried out to evaluate the prediction of the improved U-Net network. The precision P, recall R, and average cross-ratio FMIOU were used as the evaluation indexes of image segmentation, and the comprehensive evaluation index F1 was used as the evaluation value of accuracy and recall rate. The experimental results show that the improved U-Net network effectively identified and classified the complete soybean grain, broken grain, and impurities in the image. The comprehensive evaluation index values of complete, broken grain, and impurity segmentation were 95.50%, 91.88%, and 94.34%, respectively. The average intersection and MIOU were 86.83%. Correspondingly, the grain and impurity quality in the sample were determined by the impurity rate in the existing quality detection of soybean combine harvester. The crushing rate was also the ratio of broken and intact grain quality in the sample. A quantitative model was established for the broken rate and impurity rate using pixels, according to the existing measurement. Bench and field experiments were carried out using the online detection device for the soybean grain crushing rate and impurity rate. The bench test results show that the mean absolute errors were 0.13 and 0.25 percentage points for the fragmentation and impurity rate between the test and the manual, respectively. The field experiment showed that the mean absolute errors were 0.18 and 0.10 percentage points for the fragmentation and impurity rate between the test and the manual, respectively. Therefore, the proposed detection can be expected to accurately online estimate the crushing rate and impurity rate of mechanically harvested soybean. The finding can provide technical support for the online detection of the quality of soybean combined harvesting.