Customized Semantic Segmentation for Enhanced Disease Detection of Maize Leaf Images

Abstract

Maize leaf images are affected by various diseases. Though many image processing techniques are available to identify diseased segment of a diseased maize leaf image proper methodology to segment every chunk in the leaf as disease, shadow, healthy and background using a single methodology is still in search of. So, a single line of attack is availed using Semantic Segmentation for diseased maize Leaf images through which every pixel in an image is equated to a class. Initially multiple classes in the maize leaf images are Labeled and trained. ImagedataStore and PixelLabelDatastore are used to distinguish original images and trained images. With different classes defined and trained using the Semanticseg model and later applying semantic segmentation to the diseased maize leaf images they are segmented into various classes such as healthy, diseased, shadow and background. The shadows and background are difficult to handle and with this segmentation the exact pixel count of various classes are displayed. The output of semantic segmentation is a maize Leaf image where each pixel is equated to a particular class whereas in normal CNN the output is not an image but a class.