In order to quickly and accurately detect the quality of soybean seeds by the method of seed images recognition, a method of image screening and recognition based on convolution neural network is proposed, taking the classification of normal and abnormal quality seeds as an example. The data set of soybean seed quality was established, and convolution neural network was designed to extract the image features of soybean seed. In order to improve the classification accuracy and real-time performance, the convolution neural network was optimized from the aspects of design and selection of convolution neural network structure, reduction of over fitting, acceleration of training convergence speed, and enhancement of network robustness. Finally, the 6-layer convolution neural network with 4 convolution layers, 4 pooling layers and 2 fully connected layers were selected, L2 regularization and mini batch training methods were used for the network′s optimization training and test. Comparing the results with the traditional machine learning classification methods, the experimental results show that the accuracy of the optimized convolution neural network is 98.8%, and the average detection time of a single soybean seed image is 2.96 ms, which can provide an important reference for soybean seeds quality classification.
Convolution neural network