Unsupervised Controllable Enhancement of Underwater Images Based on Multi-Domain Attribute Representation Disentanglement

The unsupervised enhancement technology for underwater images is mainly oriented towards specific distortion factors and exhibits limited adaptability towards various underwater distorted images. The content attribute(structure) of the image will migrate and change with the style attribute(appearance), resulting in an uncontrolled enhancement effect and affecting the stability and accuracy of subsequent environmental perception and processing. To address this issue, an unsupervised controllable enhancement method of underwater images based on multi-domain attribute representation disentanglement(MARD) was proposed in the paper. First, a framework of multi-domain unified representation disentanglement cycle-consistent adversarial translations was designed, thereby enhancing the algorithm’s adaptability to multiple distortion factors. Subsequently, a dual-encoding and conditional decoding network structure was constructed. Finally, a series of losses for MARD was designed to enhance the independence and controllability of quality, content, style, and other attribute representations. Experimental results demonstrate that the proposed algorithm not only eliminates various distortions such as color aberration, blur, noise, and low illumination in underwater images but also quantify the image style codes by linear interpolation for controllable enhancement of underwater images.