Research On Edible Fungus Fruit Body Diseases Identification Based On Lightweight CNN

The edible fungus industry is the fifth largest industry next to grain,oil,vegetable and fruit.Developing edible fungus industry will be an important way to solve the future human food source.The occurrence of edible mushroom disease will seriously affect the yield and quality of edible mushroom,which is very important for the early identification of the disease.Crop disease recognition based on convolutional neural network is a commonly used method at present.The traditional convolutional neural network model requires a large amount of calculation and high hardware configuration requirements,so it is difficult to meet the requirements of practical application.In order to solve the above problems,this paper adopts the lightweight convolutional neural network model and takes the fruition bodies of healthy edible fungi,physiological diseases of edible fungi,bacterial diseases and fungal diseases as the identification objects,and carries out the following main work:(1)Establish the data set of edible fungus disease.Since August 2022,edible fungus fruiting body images have been continuously collected,and a total of 1135 original images have been collected.Data enhancement preprocessing methods such as noise enhancement,noise reduction,horizontal flip and vertical flip to increase light intensity have been used to construct edible fungus fruiting body disease data set.(2)The four lightweight CNN models such as MobileNet V2,DenseNet,EfficientNet and ShuffleNetv2 are selected in this paper,aiming at the cumbersome workflow and low recognition rate of manual detection methods,the long training time,large number of weight parameters and weak generalization ability of traditional convolutional neural networks.The recognition rate was 85.72%,88.50%,89.29% and 91.34%,respectively.Compared with traditional deep learning methods,the recognition rate was significantly improved,showing advantages such as small memory occupation,short recognition time and fast training speed,which made up for the shortcomings of traditional deep learning methods.(3)In order to improve the feature extraction capability of lightweight convolutional networks,the method of adding channel attention mechanism and optimizing model network structure was adopted to improve the fruiting body disease recognition network of edible fungi.Firstly,the SE module is deeply integrated with the network to make the network pay more attention to the target region and improve the model disease classification performance.The recognition rates of MobileNet V3,DenseNet+SE,EfficientNet+SE and ShuffleNetV2+SE models were 91.72%,92.50%,92.17% and 93.34%,respectively.ShuffleNetV2 was the best model.Then,the optimal model ShuffleNetV2 is improved.In order to simplify the convolution operation,the 1×1 convolution layer after the 3×3 deep convolution is deleted,and the improved network Shufflenetv2-Lite is obtained.The recognition rate of its model is 93.89%.In order to further improve the model recognition rate,the SE module was added to the improved network ShuffleNetV2-Lite,and the ShuffleNetV2-Lite+SE model was obtained,and the model recognition rate was 95.76%.Experimental studies have shown that lightweight convolutional neural networks have excellent feature extraction capabilities and can be applied to the recognition of edible mushroom fruiting entity diseases.ShuffleNetV2’s addition of attention mechanism and optimization of model network structure can effectively improve the accuracy of model recognition.This method can provide a technical framework for image processing and classification recognition of edible mushroom diseases.