Crop Disease Image Recognition And Model Compression Based On Deep Learning

Crop diseases are an important cause of crop yield decline,enabling timely and accurate identification of crop disease types is of great importance to protect food production.Traditional disease identification methods rely on human eye observation and are very prone to misjudgment.Deep learning techniques that have emerged in recent years have good performance in image recognition,but there are three challenges when applied to crop disease image recognition: Firstly,the robustness of the model is weak,and the accuracy of the model is greatly reduced when the trained model is applied to the real environment.Secondly,the accuracy of the model is low,which cannot meet the accuracy requirements of agriculture.Thirdly,the high complexity of the model makes it difficult to deploy on agricultural platforms with limited computing and memory resources.To address the above problems,this thesis carries out a research on crop disease image recognition based on deep learning as follows.(1)To address the problems of weak model robustness and low accuracy under complex natural conditions,this thesis proposes a disease recognition method based on image segmentation and convolutional neural network.The method firstly uses the traditional image segmentation algorithm to process the background of the image as black in the data pre-processing stage,so that the training set images and the images under complex natural conditions have the same background,thus reducing the influence of the complex image background on the robustness of the model.Then,the Coordinate attention mechanism is improved and embedded into the existing convolutional neural network to improve the accuracy of the model.(2)To address the problem of high model complexity that is difficult to deploy to resource-limited platforms,this thesis proposes a model compression method based on channel pruning and lightweight convolutional neural networks.The method is based on a lightweight network and embeds an attention module in the model to improve the performance.Then,the trained lightweight model is channel pruned with minimum loss of accuracy to further reduce the size,computation and response time of the model,so that the complex model that could not be used originally due to resource constraints can be compressed and applied to resource-limited platforms.