SAR Image Colorization for Comprehensive Insight Using Deep Learning Model

Abstract

Image colorization is a long-standing problem in computer vision that aims to generate visually plausible color information from grayscale images. Conventional colorization methods require manual work because they use heuristic methods which create problems when applying these methods to different types of images. The paper introduces a deep learning system for SAR image colorization which uses a pre-trained Caffe convolutional neural network to extract chrominance data from monochrome SAR images. The proposed approach works in CIELAB color space because the model needs the SAR luminance (L) channel to estimate the a and b chrominance channels. The system generates the final colorized output by merging the predicted chrominance components with the original luminance channel and transforming the result into RGB color space. The system implements a full deployment pipeline which handles image preprocessing, neural network inference, colorspace transformation, and post-processing to improve output quality. The system operates within a lightweight application which enables users to upload images, see results instantly, and download results. The experimental results show that the framework successfully improves the visual understanding of SAR images while keeping the structural elements intact. The upcoming research will focus on three main areas which include domain-specific fine-tuning, the creation of new SAR datasets, and the development of better methods for evaluating visual quality.