Abstract: In order to improve the rapid and accurate recognition of grains and impurities in direct corn harvest, a single-layer recognition stepwise classification method based on machine vision is proposed. Watershed segmentation is performed on the collected single-layer original image, and the area that is still stuck is processed by R+G color grayscale processing to construct a grayscale gradient and mark the background, and then perform a second watershed segmentation. A total of 27 feature information including color, shape, and texture of corn kernels and impurities were extracted, and then stepwise discriminant analysis was performed to obtain 13 main feature parameters that can distinguish between kernels and impurities. Based on the composite feature threshold setting, the first impurity is screened and then the ear core and stem and leaf impurity types are identified. The comparison analysis is based on the BP neural network and the genetic algorithm optimized BP neural network（Gabp neural network） to identify the integrity of the corn kernel. Compared with the bp neural network, the accuracy rate of the gabp-based neural network is increased by 1% and the experimental results are more stable. The neural network training and recognition after feature optimization takes less time. The step-by-step classification and identification method was used to identify the mixed direct-harvest corn kernels. The experiment showed that the step-by-step recognition accuracy rates were 97.7% and 85.63%, and the recall rates of intact corn kernels, broken corn kernels and impurities were 89.31%, 71.43% and 81.82%, respectively. The stepwise classification and identification method of corn kernels and impurities proposed in this paper can quickly and accurately distinguish impurities, intact corn kernels and broken kernels, which can provide reference for the design and research of the online monitoring system for direct harvest of corn kernels and broken kernels.