Maize Leaf Spot Disease Recognition Experiment Based on Deep Learning Object Decetion

China is a large maize producing country in the world. During its planting period, leaf spot is a common but serious disease and insect pest, which will reduce maize yield and damage maize quality. The use of target detection and recognition model for pest detection has gradually become a trend with developing deep learning. To quickly and accurately identify maize leaf spot disease and insect pests, keras-yolov3 model for target detection was used in this paper. This model add a layer of residual network, which not only ensured the recognition accuracy, but also increased the recognition accuracy. The multi-scale feature extraction algorithm was adopted, which can improve the detection accuracy, optimize the detection effect and improve the detection accuracy under the same parameters. In the training stage, the optimal epoch setting value was determined by comparing the degree of model fitting under different iterations. It was found that when epoch was set to 250, the training effect was the best by comparing loss curves. In the testing phase, multiple confidence settings were set and the results were verified and compared. it was found that when the confidence was 0.5, the comprehensive evaluation F1 was 0.913, the accuracy was 94%, and the recall rate was 88%, which is the best result. To ensure the real-time performance of detection, the model was tested with the same data set on CPU and GPU. The results showed that the detection speed of CPU was much lower than that of GPU. The average detection time of single image in GPU was 56ms, which was more in line with the scene of real-time target detection.