CNN Pruning for Edge Computing-Based Corn Disease Detection with a Novel NG-Mean Accuracy Loss Optimization

Abstract

Plant disease detection studies disease attacks in plants detected on the leaves using computer vision. However, some plant disease detection solutions still utilize cloud computing, where the problems include slow processing times and misuse of data privacy. This study aims to evaluate the performance of convolutional neural network (CNN) pruning in edge computing-based plant disease detection. We use Kaggle's plant disease image dataset, which contains three corn diseases. We also created an edge computing system architecture for plant disease detection utilizing the latest communication technology and middleware. Next, we developed an optimal CNN model for plant disease detection using grid search. We pruned the CNN model in the final step and tested its performance. In this step, we developed a novel normalized-geometric mean (NG-mean) method for accuracy loss optimization. The test results show that class weights can optimize specificity and g-mean on the imbalanced dataset, with values of 0.995 and 0.983, respectively. The grid search results then optimize the optimization method's hyperparameters, learning rate, batch size, and epoch to achieve the highest accuracy of 0.947. Applying pruning produces several models with variations in sparsity and scheduling methods. We used the new NG-mean method to find the best compressed model. It had constant scheduling, 0.8 sparsity, a mean accuracy loss of 1.05%, and a CR of 2.71×. This study enhances the efficiency and privacy of plant disease detection by utilizing edge computing and optimizing CNN models, leading to faster processing and better data security. Future work could explore the application of the novel NG-Mean method in other domains and the integration of additional plant species and diseases into the detection system.