Towards Achieving Lightweight Deep Neural Network for Precision Agriculture with Maize Disease Detection

Abstract

Agriculture is the pillar industry of human survival. However, various crop diseases reduce the human food supply and lead to starvation and death in the worst cases. Experts perform visual symptoms observation for crop disease diagnosis. Which process is time-consuming and expensive. Also, the process has significant risk of human error due to subjective perception. Convolutional Neural Networks (CNN) use image processing techniques to show great potential in plant disease detection. However, it requires thousands of channels to learn rich features, resulting in large models requiring powerful computing, power supply, and high bandwidth, making it more expensive and difficult for farmers to acquire. Therefore, deploying these solutions on resource-constrained devices is desirable to make them more accessible. Thus, we propose a lightweight object detection CNN that can run on resource-constrained devices to detect crop diseases. Channel pruning is applied to optimize resource use by removing unimportant channels and filter weights to reduce network parameters, inference time, and the number of FLOPS. Experimental results with object detector, Faster R-CNN with two backbones, ResNet-50, and EfficientNet-B7, show significant improvement in model efficiency, keeping high accuracy.