Leaf Image-Based Plant Disease Diagnosing CNN Compression Using Network Pruning and Weight Clustering

Abstract

AbstractDeep convolutional neural network (CNN) is proved to be efficient for diagnosis of plant disease automatically using leaf images. But the memory usage and speed of successful CNN models are often discouraging to be used in real time or to run on edge devices. So, there are numerous studies to compress and speed-up CNNs. In this paper, two memory usage effective compression algorithms (network pruning and weight clustering) are used on a trained CNN that can classify plant disease with 98.882% accuracy. The main CNN’s size is 15.93 MB. Using the pruning and clustering compression techniques, this study has succeeded to make a 4.54 MB model with 98.882% accuracy and 1.25 MB model with 98.882% accuracy. In between the algorithms, K-means weight clustering-based compression proved to be highly effective in this case. The model is compressed without any loss of accuracy over the 13 classes, and the speed of the model execution increases more than twice. The dataset that is used is a subset of PlantVillage dataset with normal and diseased potato and tomato leaves. For using CNNs real time and efficient way, robustness is a limitation; so efficient compressed model is a need of the hour. Plant disease prediction is already well discussed, but controlling the robustness and maintaining the accuracy is not studied before—that is, the main motivation of this study.KeywordsConvolutional neural network (CNN)Deep learningMachine learning in agricultureModel compressionPruning