PRECISE IDENTIFICATION OF DISEASED LEAF REGIONS USING DEEP LEARNING- BASED SEMANTIC SEGMENTATION TECHNIQUES

Abstract

Plant leaf disease segmentation has been a very important topic for research over a decade. Segmentation helps to calculate the spread of infection, and proper image preprocessing leads to better results in disease detection and identification. In this thesis we have proposed our work, findings and potential of AI models in segmenting image accurately based on recently proposed architectures. Making models too complex makes it computationally expensive. We must focus on finding the model that has less trainable parameters. This is achieved when we focus more on making architecture lossless. It can be achieved in many ways, we have proposed propagating the learnings of each previous layer to propagate in subsequent layer to avoid any information loss and up sample using indices to reduce loss. Model maintains good accuracy with very a smaller number of parameters. Our model uses a few dense layers to trap the information within the encoder decoder architecture. This reduces exponential increment in number of channels and stills make network deep and more lossless.