Mobile-oriented Plant Disease Image Recognition Algorithm Research

Currently,plant disease image recognition on mobile devices has become one of the key tools to improve the efficiency of agricultural production.However,deep convolutional neural network models are difficult to deploy and recognise on mobile devices where computational power and storage space are limited.To achieve efficient plant disease image recognition on resource-constrained mobile devices,it is particularly important to explore ways to reduce model size and computational effort.In this paper,we optimise the model for Res Net50 and Shuffle Net V2 networks and compress the model using a knowledge distillation algorithm.By deploying the optimised model to Android,we achieve the goal of plant disease image recognition on mobile.The main research of this paper includes:1.To address the issue of plant disease recognition on mobile devices with limited computational capacity and storage space,a plant disease recognition method based on the knowledge distillation algorithm is proposed.In this process,Res Net50 is chosen as the teacher network for knowledge distillation,and Shuffle Net V2 as the student network.The teacher network introduces PSA attention to enhance feature extraction capabilities.Experimental results show that the recognition accuracy of the optimized KD-Shuffle Net V2 network reaches 97.43%,slightly higher than the lightweight Shuffle Net V2 network without distillation,while the model size is reduced by 49.43%.This proves that by improving the teacher network’s knowledge distillation method,the model can be significantly reduced while maintaining high accuracy,which is beneficial for improving the computation and storage of the model on mobile devices.2.To enhance the ability of the student network model to recognize plant disease images under complex natural conditions,we propose an improved student network model,Shuffle Net V2＿res2net＿eca.This model introduces Res2 Net structural blocks and the ECA attention mechanism into the original Shuffle Net V2 network,thereby enhancing the extraction of multi-scale image features and the recognition of plant disease spot features.Experimental results indicate that the recognition rate of the student network Shuffle Net V2＿res2net＿eca optimized by knowledge distillation reaches 86.21% on the apple disease dataset in a complex natural environment,an improvement of 6.25% compared to before optimization.Therefore,the model demonstrates high accuracy and robustness in handling plant disease image recognition tasks in complex environments.3.To verify the performance of the distilled network model on mobile devices,an Android-based application is designed and implemented to intuitively demonstrate the plant disease recognition effect of the model on mobile devices.By employing the knowledge distillation method to compress the model size to a level suitable for mobile devices,and deploying the trained network model offline on the Android platform,an Android application with camera photography,image selection,and model inference functions is developed.Experiments in actual plant disease detection tasks confirm the effectiveness of the proposed method in plant disease recognition,demonstrating its practicality and feasibility.