Research On Maize Quality Recognition Based On Generative Adversarial Network And Transfer Learning

As the most fundamental production material in agriculture,the quality of seeds largely determines the harvest results and directly affects farmers’ income,which is of great significance for the steady development of the rural economy.Henan Province is one of the major corn producing areas in China,with a corn planting area of approximately 50 million mu and leading corn breeding capabilities in the country.The quality of corn seeds is of great importance for the price,storage,transportation,and industrialization of corn seeds.With the rapid development of deep learning technology,many new methods based on it have emerged for corn quality recognition research.However,due to the lack of data and the problem of sample imbalance,the generalization ability and recognition robustness of deep learning models have been affected.Based on this,this paper proposes a new method for corn quality recognition under the conditions of data sample missing and sample imbalance,by using generative adversarial networks(GAN)and its improved M_RWS_DCGAN(Maize_RWS_DCGAN)for corn image data enhancement and transfer learning for recognition.Firstly,a sample collection platform is built for image acquisition.Corn seeds are double-sided photographed to obtain as many samples as possible,and a total of 1678 images(1200 normal grains,202 damaged grains,148 moldy grains,and 128 shrunken grains)are collected as the original dataset.Secondly,based on GAN,the dataset is optimized and enhanced using DCGAN and the improved M_RWS_DCGAN,generating high-quality corn images and supplementing the corn quality dataset,thus solving the problem of data sample imbalance.In addition,for corn quality recognition,a transfer learning approach and feature fusion method are used to replace the classifier in the fully connected layer(FC layer)using the DenseNet121,ResNet50,and VGG16 pre-trained models.Traditional image features such as gray-level co-occurrence matrix(GLCM)and histogram of oriented gradients(HOG)are also incorporated and fused with deep features extracted by the pre-trained models,effectively improving the recognition accuracy of the model.Finally,the paper presents the verification of the method.The experimental results show that M_RWS_DCGAN can efficiently generate high-quality corn images,and the average accuracy of corn quality recognition based on transfer learning is 87.36%.After feature fusion,the average recognition rate for corn quality reaches 92.29%(with recognition accuracies of 91.67%,93.75%,92.08%,and 91.67%for normal grains,damaged grains,moldy grains,and shrunken grains,respectively).The research in this paper provides an effective and convenient method for data set construction and recognition in agricultural product detection,which is helpful for improving the quality of corn planting and effectively managing the post-harvest circulation of corn in Henan Province and even in China.