Research And Implementation Of Mobile-oriented Multi-class Crop Disease Recognition Method

In recent years,the rapid advancements in deep learning theory and computational hardware capabilities have brought about revolutionary changes in various industries.Utilizing Convolutional Neural Networks(CNN)to address the problem of identfying crop disease and pest issues has gradually become a focal point of interest.Smartphones have progressively integrated into people’s daily lives,and using smartphone terminals as a medium for image recognition and developing applications with disease and pest identification capabilities has emerged as a new research direction.Therefore,this study focuses on this particular research direction,with the main content as follows:(1)In crop pest identification tasks,it is challenging to obtain a sufficient number of samples.To this end,this paper presents a leaf image translation model based on an improved CycleConsistent Adversarial Networks(CycleGAN).The model achieves data enhancement of a limited number of diseased leaf images by combining healthy leaves with a small number of diseased samples for unsupervised image translation.To generate more realistic diseased leaf images,the Class Activation Maps(CAM)attention mechanism is introduced to guide the generator to focus more on the target leaf image features in response to the inability of the original CycleGAN structure to distinguish the most distinctive semantic features of the images.To address the blurred texture and distorted hue of the generated leaf,Texture Reconstruction Loss(TRL)is introduced to measure the difference in texture between the generated disease image and the target image,so that the generator can better preserve the texture details of the leaf during the learning process.The experimental results show that the improved CycleGAN generates higher quality leaves and the generated disease samples can effectively improve the disease recognition accuracy.(2)To address the problem of a single background of disease disclosure data,the training model is difficult to apply to the field environment for identification.In this paper,a complex environment dataset is constructed and the CE-MobileNet V2(Complex Environment MobileNet V2)network structure is proposed to address the problem of low performance of the lightweight network MobileNet V2 in complex environments.Firstly,Coordinate Attention(CA)mechanism is introduced to enhance the focus of the model on the spot region and reduce the influence of irrelevant background.Secondly,Multiscale Feature Fusion(MFF)is added to effectively extract features of different sizes and morphologies of the spot areas.Next,Asymmmetric Convolution Blocks(ACB)are used for alternative training to improve the feature representation capability of the convolution kernel.Finally,the number of parameters of the model was reduced by adjusting the number of network channels to achieve a balance between performance and model parameters.The experimental results show that the CE-MobileNet V2 network structure is able to achieve high recognition accuracy in a field environment while having small model parameters.(3)An integrated foliar disease identification system has been designed and implemented with four functions: registration and login for personal information registration and system login;environmental monitoring for real-time access to environmental parameters in the field,such as temperature,humidity and light;The plant disease identification function uses the CE-MobileNet V2 model proposed in Chapter 4to achieve fast and accurate disease identification on local mobile devices;the disease encyclopaedia function provides detailed information on various foliar diseases.The system meets the practical needs of growers and provide them with convenience and applicability.