What question did this study set out to answer?

The aim is to create a pipeline that generates infrared images from standard RGB inputs, overcoming high costs of infrared cameras.

April 5, 2026Open Access

V2IR: Visible-to-Infrared Image and Video Translation Using Pix2Pix Generative Adversarial Network with Integrated YOLOv8 Object Detection

Key Points

The aim is to create a pipeline that generates infrared images from standard RGB inputs, overcoming high costs of infrared cameras.
Developed V2IR using a Pix2Pix Conditional Generative Adversarial Network
Utilized a U-Net architecture with 256 depth and trained on the LLVIP dataset
Implemented YOLOv8 for object detection integrated with the synthesised infrared outputs
Conducted qualitative and quantitative assessments over 200 training epochs
Delivered the system as a Flask web application for real-time image and video processing
Achieved high-quality infrared images from RGB inputs with improved thermal fidelity
Successful localization of persons and vehicles even in poor RGB conditions
Demonstrated significant enhancements in detection confidence after training
Validated as a software-based solution that minimizes the need for expensive infrared hardware

Abstract

Infrared imaging plays an indispensable role in night-vision surveillance, autonomous navigation, search-and-rescue operations, and defence systems, yet the hardware infrastructure required to capture genuine thermal data remains prohibitively expensive for large-scale deployment. This paper introduces V2IR, a complete end-to-end pipeline that synthesises perceptually convincing infrared representations from ordinary RGB photographs and video streams by means of a Pix2Pix Conditional Generative Adversarial Network. The generator adopts a U-Net 256 architecture with eight levels of skip-connected encoder–decoder blocks trained over 200 epochs on the Large-scale Visible-Infrared Paired (LLVIP) dataset comprising 15,488 co-registered image pairs. A linear learning-rate decay schedule, Adam optimisation with β1 = 0.5, and a composite loss combining vanilla GAN cross-entropy with an L1 reconstruction term weighted at λ = 100 are employed. The synthesised infrared output is subsequently fed into a YOLOv8 anchor-free detector hosted on the Roboflow Serverless API and fine-tuned on FLIR thermal imagery, achieving reliable person and vehicle localisation under conditions where RGB-domain detectors degrade severely. The complete system is delivered as a Flask web application that supports single-image inference, URL-based input, and frame-level video processing with real-time progress streaming via Server-Sent Events. Qualitative and quantitative analysis across different training epochs demonstrates progressive improvement in structural fidelity, thermal contrast, and detection confidence, validating V2IR as a cost-effective software-based alternative to dedicated infrared acquisition hardware.

Read Full Paperexternally

AIに質問

Bookmark

View Full Paper