Unsupervised Exposure Correction Achieves Detail Restoration Without Manual 2-Step Labeling

Researchers are tackling the persistent problem of exposure correction in images, a challenge often resulting in lost detail and colour distortion. Puzhen Wu, Han Weng from Beijing-Dublin International College, University College Dublin, and Quan Zheng from the Institute of Software, Chinese Academy of Sciences, et al., present a novel unsupervised approach that addresses key limitations of current techniques. Their work moves beyond simply adjusting brightness by incorporating semantic understanding of image content, effectively reducing colour shift artifacts, and crucially, does so without requiring the laborious manual labelling of training data. By leveraging pre-trained models like Fast Segment Anything and CLIP, alongside a new pseudo-ground truth generator, this research demonstrates significant improvements over existing unsupervised methods, offering a promising pathway towards more accurate and efficient image restoration.

Scientists have developed a new unsupervised method for correcting exposure in images, addressing critical challenges in image processing and computer vision. The research team tackled the issues of detail loss, colour distortion, and reduced contrast caused by improper exposure.

Semantic-fused spatial mamba for exposure correction offers improved

Scientists developed a novel unsupervised semantic-exposure correction network to address deficiencies in correcting improperly exposed images. The research tackled two key challenges: the lack of object-wise regional semantic information leading to colour shift artefacts, and the absence of labelled data requiring extensive manual editing. This fused feature set then drives a multi-scale residual spatial mamba group designed to restore image details and adjust exposure levels.

To circumvent the need for manual labelling, researchers proposed a pseudo-ground truth generator guided by CLIP, a vision-language model. This generator was fine-tuned to automatically identify exposure situations and subsequently instruct tailored gamma corrections, creating synthetic training data. Furthermore, the study harnessed rich priors from both FastSAM and CLIP to develop a semantic-prompt consistency loss. This loss function enforces semantic consistency and image-prompt alignment during unsupervised training, ensuring the corrected images align with semantic understanding. The core of the network comprises multi-scale Semantics-Informed Mamba Reconstruction (SIMR) blocks, incorporating both downsampling and upsampling operations.

Within each SIMR block, an Adaptive Semantic-Aware Fusion (ASF) module fuses semantic features with image-space features, while a Residual Spatial Mamba Group (RSMG) module refines exposure correction by capturing long-range spatial dependencies. Experiments employ this network to correct real-world exposure images, demonstrating performance gains over state-of-the-art unsupervised methods, both numerically and visually, as illustrated in Figure 0.1. Quantitative evaluation reveals improvements in both PSNR and SSIM metrics; performance comparisons show values ranging from 10 to 22 for PSNR and 0.45 to 0.85 for SSIM, depending on the specific configuration and compared method. The team’s approach introduces a novel framework consisting of a semantic-aware exposure correction network and an automated pseudo-ground truth generation process, eliminating laborious manual labelling and promoting high-quality unsupervised training through semantic and image-prompt alignment.

Semantic guidance improves unsupervised exposure correction by leveraging

Scientists have developed a new unsupervised semantic-exposure correction network to address challenges in correcting real-world images suffering from improper exposure. The research tackles two key issues: the lack of object-wise regional semantic information leading to colour shift artefacts, and the difficulty of obtaining ground-truth labels for training, which typically requires extensive manual editing. The core of the breakthrough delivers a multi-scale residual spatial mamba group designed to restore image details and refine exposure levels.

Measurements confirm that this architecture captures long-range spatial dependencies, improving the fidelity of the corrected images. To circumvent the need for manual labelling, researchers propose a pseudo-ground truth generator guided by CLIP, which was fine-tuned to automatically identify exposure situations and instruct tailored corrections. Tests prove the generator accurately classifies images, providing a valuable training signal without human intervention. Data shows the team leveraged rich priors from both FastSAM and CLIP to develop a semantic-prompt consistency loss. This loss enforces semantic consistency and image-prompt alignment during unsupervised training, ensuring that the corrected images remain semantically coherent and visually appealing.

Comprehensive experimental results illustrate the effectiveness of the method in correcting real-world exposure images, demonstrably outperforming state-of-the-art unsupervised approaches both numerically and visually. Specifically, performance comparisons, as shown in Figure 0.1, indicate improvements in both Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR). At a PSNR of 10, the method achieves an SSIM of approximately 0.45, increasing to an SSIM of 0.70 at a PSNR of 20. Similarly, at a PSNR of 7, the SSIM is around 0.60, rising to 0.85 at a PSNR of 22. These measurements confirm the network’s ability to significantly enhance image quality and restore details lost due to improper exposure. The framework consists of Semantics-Informed Mamba Reconstruction (SIMR) blocks with downsampling and upsampling operations, and an Adaptive Semantic-Aware Fusion (ASF) module.

Semantic Fusion and Spatial Mamba for Exposure offer

Scientists have developed a new unsupervised framework for exposure correction that improves visual quality and metric scores in images. The research addresses two key challenges in exposure correction: the lack of semantic understanding of image regions and the difficulty of obtaining labelled data for training. This fused information is then processed by a multi-scale residual spatial mamba group to restore image details and adjust exposure levels.

A pseudo-ground truth generator, guided by CLIP, automatically identifies exposure issues and directs corrections without requiring manual labelling. Furthermore, a semantic-prompt consistency loss leverages the capabilities of FastSAM and CLIP to ensure semantic consistency and alignment between images and prompts during unsupervised training. Extensive experiments demonstrate the effectiveness of this method in correcting real-world exposure images, surpassing the performance of existing unsupervised techniques. The authors acknowledge that inference speed decreases with very high-resolution images.

Future research will focus on utilising generative models for localised inpainting to recover details in severely over or underexposed areas. Exploration of lighter network backbones and techniques like pruning or distillation are also planned to reduce computational demands. Extending the approach to dynamic video and multi-camera systems, utilising advances in camera motion understanding and calibration, represents a further potential direction for development.