Abstract: This study was aimed to solve the problems of poor applicability and low recognition efficiency of existing unsound wheat kernels detection methods, an improved YOLOX-based wheat imperfection grain detection method was proposed. Wheat images of three varieties, different heights and number of seeds scenarios were collected to produce a dataset. Meanwhile, this study introduced coordinate attention model in YOLOX backbone network to enhance the saliency of imperfect grains in images, and the feature fusion structure is improved by using a weighted bidirectional feature pyramid to realize the effective fusion of imperfect grain features at different scales. To address the problem of real-time multi-target detection, the depth-separable convolution module was used to lighten the feature extraction network, reduce the computational power of the network, and improve the detection speed of the network. The results showed that the improved model had a detection rate of less than 5%, an average detection accuracy of 93.43%, and a detection speed of 37 fps, which were 2.58% and 5.78 fps higher than those of the original YOLOX network. The improved model can effectively identify and detect wheat imperfect grains at different number of seeds and different densities, which can provide technical reference for further real-time detection and statistical application of wheat imperfect grains.