Abstract: Accurate identification of maize seedlings is an important prerequisite for the realization of automatic and accurate weeding, thinning, replanting and other seedling operations. In this paper, based on the characteristics of CNN cascade neural network and cascade neural network,a fast and accurate detection method for corn seedling is proposed. Firstly, the feature map of corn seedling image is extracted by ResNet50 and FPN,and the regional coordinate feature is extracted by RPN（Region Proposal Network）. Then, the feature map of fixed size is obtained by Ro IAlign layer, and the output module is used to calculate the classification, regression and segmentation of the feature map to complete the calculation of the specific orientation, category and contour of crops. In this study, field images of 3-5 leaf maize seedlings were collected and data sets were made R-CNN model is trained, and AlexNet, VGG16, ResNet18, ResNet50 and ResNet50 + FPN are selected as feature extraction networks respectively for comparative experiments. Resnet50 + FPN is determined as the optimal feature extraction network, with the average accuracy（mAP） of 91. 76% and the average detection time of 6. 5ms; fast, a two-stage target detection model, is selected R-CNN, R-FCN and CoupleNet are compared with Cascade R-CNN which takes ResNet50 + FPN as feature extraction network. The results show that Cascade R-CNN has the best model detection effect and the fastest speed. Therefore, this method can effectively detect corn seedlings in the natural environment, and provide technical support for automatic precision operation of corn seedlings.