Momentum Contrast Learning for Aerial Image Segmentation and Precision Agriculture Analysis

Abstract

Computer vision in precision agriculture analysis has gained increasing attention as recent advancements in deep learning-based methods for various tasks were proven successful. As one of the primary problems in agriculture-vision applications, semantic segmentation from aerial agricultural images, differs from common object or aerial image segmentation tasks in various ways. Recently, there have been some efforts that aim to apply deep learning techniques to model multi-spectral aerial images and segment field anomaly pattern objects with extremely irregular shapes and scales. However, most existing methods fail to propose effective methods for model initialization and perform poorly in segmenting small objects. To address these challenges, we propose a deep learning framework that leverages momentum contrast learning with a PointRend-based model for aerial image analysis. Extensive experiments have demonstrated the effectiveness of our model for better aerial image semantic segmentation.