Abstract: The impurity rate is one of the important harvesting performance indexes of the rice combine harvester. The impurity of the harvested grain during the harvest includes the branches and straws of the crop. In order to study the correlation between the impurity rate of the grain and the operation parameters of the combine harvester, it is necessary to obtain the data of impurity rate in real-time. This paper studies the segmentation algorithm of hybrid rice grain image based on U-net model of machine vision technology. Aiming at the problems existing in traditional segmentation algorithm, such as large amount of computation, time-consuming processing, serious over segmentation of images, and the determination of segmentation parameters depends on human experience, the deep learning model is used to train and learn image features of each segmentation category of pixel level for many times. Based on the U-net depth learning model, a method of predicting and segmenting grains, branched and straws in hybrid rice grain images is proposed. The improved U-net network is used to increase the depth of the network and add the batch normalization layer. The information of more abundant data is obtained in a small data set, and the problem of lack of training data and over fitting of training is solved. In this paper, 50 rice images collected from the field experiments are selected, Labelme method is used to annotate and enhance data. 1 000 small samples of 256 × 256 pixels are cut, in which 700 images are used as training data set, 300 images are used as verification data set, and a hybrid rice grain image segmentation model of combine harvester based on improved U-net network is established. The accuracy of the model is measured by the comprehensive evaluation index, and 60 images with 8-bit RGB selected randomly are verified. The experimental results show that the comprehensive evaluation index value of rice grains segmentation is 99.42%, the comprehensive evaluation index value of branch and stem segmentation is 88.56%, and the comprehensive evaluation index value of straws segmentation is 86.84%. The proposed algorithm based on U-net model can effectively segment the grains, branches and straws in the hybrid rice grain images, and has the higher real-time and accuracy of the segmentation. The research results can provide technical support for the further recognition and processing of rice grain image, and provide algorithm reference for the design of rice combine harvester impurity rate monitoring system.