LRMC-DeepLabV3+: Multiclass Leaf Disease Semantic Segmentation Based On An Improved DeepLabV3+ Network

Abstract

The accurate and efficient segmentation of plant diseases is the key to sustaining plant growth quality and identifying disease severity. Unfortunately, many current plant disease segmentation techniques frequently fall short of precisely and quickly identifying diseased areas on plant leaves, more specifically when it comes to lightweight segmentation models with the goal of achieving high-level accuracy. The proposed model within this study will be using an improved version of DeepLabV3+ as the foundation for a deep-learning strategy that is intended to quickly and precisely segment common leaf diseases in six different species of plants. In order to modify and better the segmentation performance, the approach for this model combines the CBAM-FF (Convolutional Block Attention Module Feature Fusion) module which uses two analytical dimensions known as spatial attention and channel attention and they are needed to create a sequential attention structure that moves from channel to space. Moreover, the Lite R-ASPP (Lite Reduced Atrous Spatial Pyramid Pooling) attention module has been utilized to enhance the MobileNetV3_large backbone network's feature extraction performance for disease features. Furthermore, the impact of the optimizer, backbone network, and learning rate on the DeepLabV3+ network model's performance will be examined. The proposed model depicted an outstanding accuracy of 97.34% alongside an MIoU of 93.47%.