Research And Implementation Of Computer Vision-based Algorithm For Soybean Leaf Diseases

Accurate recognition of soybean diseases is a key factor in reducing the spread of soybean diseases to improve soybean yield,and soybean leaf parts are prone to diseases and have a long growth cycle.Therefore,this paper develops a soybean leaf disease recognition system using computer vision technology to provide users with accurate prediction types and visual interpretation of soybean leaf diseases,which is important for improving soybean yields.In this paper,accurate recognition of soybean leaf diseases was achieved by background segmentation,data augmentation,transfer learning,and classification algorithms for eight diseases commonly found in soybean leaves.Realistically collected soybean leaf images usually have complex backgrounds containing interference features that degrade the performance of recognition models.Therefore,this paper proposes a leaf image segmentation method based on a combination of the K-means algorithm and the maximum connected domain algorithm,which can effectively remove the complex background from soybean leaf images and crop out images containing only the leaf region,while the leaf can retain disease features.The method can avoid the influence of disturbing features in the image background on the recognition model and improve the robustness of the recognition model;it can save the acquisition cost by not needing to pay attention to the background area of the image when acquiring the image.Increasing the size of the dataset is significant to improve the recognition performance of convolutional neural networks,however,the acquisition of leaf disease images,in reality,is a timeconsuming and expensive task and requires experts to identify and label the types.Traditional data augmentation(e.g.,rotation,flip,translation,etc.)generated images are limited by processing rules to obtain new disease features,so this paper proposes a generative adversarial network(GAN)data augmentation model while compressing the model to reduce its size.The model can generate highquality soybean leaf disease images that have more variable and multivariable features,which can save the cost of manual labeling of soybean leaf diseases and improve the performance of the recognition model.The model is small and suitable for deployment to the application side.Deep learning networks are time-consuming to train and consume a lot of computational resources.Transfer learning can be trained based on pre-trained model weights in very large datasets to optimize training efficiency.However,the traditional fine-tuning method of transfer learning is to manually select the freeze and fine-tune layers,which may not apply to the target tasks in other domains.Therefore,this paper proposes an improved strategy for transfer learning,which uses a Bayesian optimization algorithm to automatically fine-tune three pre-trained models(VGG16,Res Net34,Dense Net121).This method can adaptively select the optimal hyperparameter configuration for fine-tuning for the target task without manually setting the freeze and fine-tuning layers.Compared with the traditional fine-tuning method,the method obtains better performance on all three pre-trained models for the soybean leaf disease dataset,indicating that the method has good generalization ability on different models.In this paper,two different datasets(using the same original images as the validation set)are formed using leaf disease data generated by traditional data augmentation and GAN data augmentation,and automatic fine-tuning is performed on the Bayesian-optimized pre-training models to verify that GAN data augmentation can improve the performance of the recognition models,and to verify that the transfer learning improvement strategy proposed in this paper has good generalization ability on different datasets.Setting up this controlled experiment can evaluate the optimal recognition model for application-side deployment.In this paper,an experimental system for soybean leaf disease identification is developed to deploy the above algorithm and model into a Web application that can be applied on both the computer and mobile side.A visual interpreter of leaf disease prediction results is added to the Web application in this paper to make researchers and users more confident in the results of leaf disease recognition.The significance of developing this system is that the performance of the model can be tested at the application level and the interpretability of the methods in this paper can be verified.