Abstract
RestoreX-AI introduces a task-oriented framework for image restoration that prioritizes machine perception over human aesthetics. By leveraging Explainable AI (XAI) and object detection scores to guide the training of Generative Adversarial Networks (GANs) and Transformers, the framework achieves a 178% average increase in mean Average Precision (mAP) for object detection in extreme weather conditions like heavy snow and sandstorms.
Introduction
In autonomous systems such as self-driving cars and drones, visual clarity is essential for safety. However, adverse weather conditions—fog, rain, snow, and dust—frequently corrupt visual data, rendering standard object detection (OD) models unreliable. Traditionally, image restoration models (e.g., GANs or Transformers) are evaluated using human-centric metrics like Peak Signal-to-Noise Ratio (PSNR) or Structural Similarity Index (SSIM). These metrics measure aesthetic quality rather than algorithmic utility. RestoreX-AI addresses this fundamental gap between human and machine perception by prioritizing features necessary for computer vision tasks.
Methodology
The RestoreX-AI framework employs a contrastive approach to evaluate and guide restoration models based on their performance on downstream object detection tasks.
Restoration Architectures
The study utilizes two primary architectures:
- Conditional GANs (cGANs): Specifically, Weather-NightGAN (night-to-day transitions) and Weather-RainGAN (rain-to-clear transitions).
- Restormer: A high-efficiency transformer designed for image restoration, tested across various noise levels.
XAI-Guided Training
A core innovation of RestoreX-AI is the integration of Grad-CAM (Gradient-weighted Class Activation Mapping). Instead of relying solely on final detection scores, the system generates attention maps to visualize which pixels the object detector prioritizes.
Training progress is monitored via a contrastive evaluation metric ($\Delta\phi$), which calculates a weighted sum of label similarity and "explanation probability" (the confidence score from the Grad-CAM model). This allows the system to stop training when the image reaches peak utility for detection, even if it is not the most visually "clean" version.
Experimental Results
The framework was tested using the DAWN Dataset, which contains real-world adverse weather imagery. The results demonstrate a significant improvement in machine perception across multiple categories.
Quantitative Performance
The proposed approach achieved an average 178% increase in mAP across adverse weather conditions compared to unrestored baseline images. Notably, in specific training stages of the Weather-RainGAN, motorcycle detection accuracy jumped from 0 AP to 76 AP.
The "Denoising Paradox"
An important finding of this research is that aggressive denoising does not always yield better detection. In tests with the Restormer, increasing noise levels occasionally improved detection for certain classes, such as pedestrians. This suggests that standard smoothing processes can inadvertently erase critical edge features required by detection algorithms.
Conclusion
RestoreX-AI demonstrates that image restoration for autonomous systems should be guided by task-specific utility rather than human visual preference. By using XAI as a training guide, the framework successfully bridges the gap between image reconstruction and object detection. Future work will focus on stabilizing GAN training and developing more robust explainability frameworks to further couple restoration with diverse computer vision architectures.
Summary updated