Deep Learning Methods for Plant Image Segmentation and Classification

Precision agriculture relies heavily on the crucial components of plant image segmen- tation and classification. The application of image classification is particularly related to disease identification and plant recognition, contributing to heightened accuracy and operational efficiency. Concurrently, image segmentation plays a pivotal role in extracting plant objects, facilitating yield prediction, disease localization, and weed detection.
The thesis starts with the development of innovative deep learning algorithms for autonomous precision agriculture. A novel framework for imbalanced semantic segmen- tation is proposed in Chapter 3, based on fully convolutional network architecture, a feature learning of weight update approach and an effective data balance scheme. Apart from the dynamic weight updates, learning holistic feature knowledge emerges as a piv- otal factor in enhancing overall performance. Chapter 4 introduces a novel learning network based on the Squeeze and Excitation Network, specifically designed for fine- grained plant pathology classification. This architecture integrates label knowledge and feature knowledge to represent plant diseases, surpassing the capabilities of single learning networks. It excels in the self-distillation of additional feature knowledge, addressing potential losses after multiple convolutional layers. In Chapter 5, a novel dataset and a semi-supervised annotation method are proposed, leveraging the faster region-based convolutional neural network. A deep learning architecture for seman- tic segmentation is developed to navigate challenges posed by complex backgrounds, demonstrating efficacy in both practical scenarios and benchmark datasets.
All proposed methodologies undergo testing on benchmark datasets across diverse environments, affirming their capacity for precise plant segmentation and classification.