Naturalness Quality Evaluation for Infrared Thermal Image Colorization

Objective Infrared image colorization technology aims to enhance the visual interpretability and information transmission efficiency of thermal imagery. This technology has significant applications in security monitoring, autonomous driving, and medical diagnosis. However, colorization processes may introduce distortions that affect the practical value of images. Therefore, developing objective and accurate quality assessment methods for colorized infrared imagery is critically important for ensuring visual quality and information preservation in practical applications. Methods This paper systematically reviews the research status and development trends of quality assessment methods for infrared image colorization. We analyze specific factors affecting infrared image colorization quality from multiple dimensions, including color fidelity, thermal information integrity, visual reasonableness, and artifact suppression. Existing quality assessment methods are classified and compared, focusing on traditional full-reference metrics [peak signal-to-noise ratio (PSNR), structural similarity index (SSIM)], color-related metrics (CIEDE2000, color deviation), and no-reference metrics [blind/referenceless image spatial quality evaluator (BRISQUE), naturalness image quality evaluator (NIQE), neural image assessment (NIMA)]. Additionally, we propose a novel no-reference quality assessment model called LDANet that integrates deep learning with color space modeling techniques. The model employs a dual-attention mechanism combining channel and spatial attention modules embedded in an Inception network backbone for adaptive multi-scale feature fusion. It incorporates color quantization, K-means clustering, and latent Dirichlet allocation (LDA) modeling to capture color distribution characteristics efficiently. Results and Discussions To evaluate the performance of various quality assessment methods, we construct a subjective evaluation database using multiple colorization algorithms [including CycleGAN, TICCGAN, TGCNet, TIVNet, and region-based self-segmentation guided diffusion model (RSDM)] applied to images from public datasets (FLIR, KAIST) and our custom-built urban infrared image dataset. Comprehensive experiments demonstrate that deep learning-based methods, particularly our proposed LDANet, outperform traditional metrics in terms of consistency with human subjective perception. LDANet achieves superior performance with Pearson linear correlation coefficient (PLCC) of 0.915, Spearman rank-order correlation coefficient (SROCC) of 0.920, and root mean squared error (RMSE) of 0.105 (Table 2), significantly better than conventional approaches. Ablation studies (Table 3) confirm the effectiveness of each component in our model architecture, with the LDA color topic modeling and attention mechanisms contributing substantially to performance improvements. These experimental results indicate that traditional methods are limited by the expressive capability of hand-crafted features, while learning-based approaches can better capture the complex relationship between image characteristics and human visual perception. Conclusions Despite the progress made in colorized infrared image quality assessment, several research directions remain to be explored. Developing more robust and perceptually accurate no-reference image quality assessment metrics that better capture specific quality attributes associated with colorized thermal infrared data remains an urgent need. Future research should focus on standardized, publicly available colorized infrared image datasets with subjective quality annotations to facilitate the training and objective evaluation of new quality assessment models. Additionally, further investigation into quality assessment methods with particular sensitivity to distortions and artifacts in colorized infrared images (such as color leakage and false contours) is essential. For practical applications like autonomous driving and real-time monitoring, efficient and computationally lightweight quality assessment methods have significant practical value. Addressing these challenges is crucial for fully realizing the potential of colorized infrared imaging and ensuring its reliable application across various domains.