Fieldscale: Locality-Aware Field-based Adaptive Rescaling for Thermal Infrared Image

TL;DR

Fieldscale tackles the challenge of rescaling 14-bit RAW TIR images to 8-bit by introducing locality-aware 2D scalar fields that encode both pixel intensity and spatial context. The method constructs two fields via grid min/max pooling, Local Extrema Suppression, and iterative Message Passing, followed by bilinear interpolation and contrast enhancement, enabling adaptive per-pixel gains that preserve local radiometry and details. Across IQA, visual place recognition, and object detection benchmarks, Fieldscale provides superior or competitive performance with demonstrated real-time capability and robustness to parameter choices. By releasing open-source code, Fieldscale offers a scalable, hardware-agnostic rescaling framework that improves downstream perception in safety-critical robotics tasks, with potential extensions to temporal consistency in video streams.

Abstract

Thermal infrared (TIR) cameras are emerging as promising sensors in safety-related fields due to their robustness against external illumination. However, RAW TIR image has 14 bits of pixel depth and needs to be rescaled into 8 bits for general applications. Previous works utilize a global 1D look-up table to compute pixel-wise gain solely based on its intensity, which degrades image quality by failing to consider the local nature of the heat. We propose Fieldscale, a rescaling based on locality-aware 2D fields where both the intensity value and spatial context of each pixel within an image are embedded. It can adaptively determine the pixel gain for each region and produce spatially consistent 8-bit rescaled images with minimal information loss and high visibility. Consistent performance improvement on image quality assessment and two other downstream tasks support the effectiveness and usability of Fieldscale. All the codes are publicly opened to facilitate research advancements in this field. https://github.com/hyeonjaegil/fieldscale

Figures (8)

• Figure 1: Due to the locality of the heat source, rescaling the thermal image requires a spatially adaptive strategy. In contrast to previous 1D LUT-based methods, Fieldscale can successfully handle local heat sources and enhance both global and local details.
• Figure 2: Overview of Fieldscale and examples of rescaled images with corresponding fields. By passing the RAW image into three sequential modules (Grid-wise Min/Max Pooling, Local Extrema Suppression, and Message Passing), min/max fields with the same size of the image are constructed from the local information of each area. A darker pixel indicates the extreme values: a lower value for the min field and a higher value for the max field. After applying optional image enhancement in parallel, the RAW image is rescaled with two fields. As visualized in three examples, cold regions bring down the corresponding area of the min field, and hot regions raise that of the max field. This locally adaptive field construction can successfully rescale images with high spatial temperature variance without loss of original details.
• Figure 3: A hot object only influences the corresponding area of the max field while leaving the min field unaffected. The local neighborhood information solely determines the pixel gain during the rescaling process.
• Figure 4: Qualitative comparison of rescaled images.
• Figure 5: IQA heatmap of different MP iterations and LES thresholds.