Visible-to-Infrared Image Translation for Matching Tasks

Abstract

Visible-to-infrared image translation is an important way to enrich infrared data. However, the reliability of the data generated by image translation in downstream tasks has always been controversial. This article proposes a method that integrates visible-to-infrared image translation tasks and multimodal template matching tasks. The image generation network is based on a generative adversarial networks (GANs), and network training is supervised by L1 loss, GANs loss, and match loss, where the matching loss includes normalized cross-correlation (NCC) loss and match patch (MP) loss. NCC loss is constructed based on the NCC matching algorithm. MP loss is calculated by modeling template matching as a contrastive learning problem. In experiments on the KAIST, VEDAI, and AVIID datasets, this method outperforms state-of-the-art methods in terms of image generation quality and template matching accuracy. Our method incorporates the image matching process into image-to-image translation, demonstrating the usability of GANs-based image generation for critical downstream tasks. This research resolves the practical controversy of generating images based on GANs and provides a theoretical reference for image generation for tasks, such as multisource image object detection and data association.