AIF Image Restoration through Multi-Task Learning-based Depth and Blur Estimation

Predicting defocus blur and depth from a single RGB image is an important task in computer vision applications such as scene understanding, image processing, and autonomous driving. In this paper, we propose a Multi-Task Learning (MTL) framework that performs both tasks simultaneously by leveraging the physical interaction of defocus blur and depth map. The proposed model extracts common visual features through a shared encoder, performs defocus blur and depth estimation using task-specific decoders, and strengthens the sharing of visual cues between tasks. It provides the restoration result of the AIF image through the estimated depth and blur results. To ensure the spatial consistency and robustness of the model, the spatial smoothness, blur-depth consistency, and weight normalization conditions are incorporated into the loss function. The proposed MTL framework demonstrated superior performance Mean Absolute Error (MAE), Structural Similarity In Experiments (SSIM) using synthetic and real datasets compared to Single Task Learning(STL) and existing MTL approaches. Removal studies confirmed that the loss function components contribute to improved prediction accuracy. This study provides an efficient and practical solution for joint estimation of defocus blur and depth.