YOLO26 Backbone-Based Image Deraining Model Released

Experimental results transferring the CSPDarknet backbone and PAN-FPN neck weights learned by YOLO26's depth estimation model to image deraining tasks have been released. Since depth estimation is a dense regression task more similar to image restoration than detection, this validates that the architecture can effectively transfer to other dense regression tasks.

Architecture Changes

The existing depth head was removed and a new RGBHead was introduced. Because image deraining requires pixel-level accuracy, a reconstruction tail was added to restore up to the input resolution, rather than stopping at the 1/4 resolution where the depth model stopped. Additionally, skip connections were applied from the stride-2 and stride-4 backbone layers to the tail to prevent loss of fine details, and the output was designed as a residual method that adds correction values to the original image.

Performance and Deployment

Training and evaluation were conducted using the mixed synthetic/real rain dataset from the ClearView library and the Charbonnier loss function. The released models come in two scales: nano (5.25M parameters) and small (12.13M parameters). When loading the YOLO26-depth checkpoint, 468 tensors in the backbone and neck match exactly, suggesting that the pre-trained weights remain effective for the deraining task.