A multi-scale skip feature fusion underwater image enhancement method based on cycle adversarial network

To address the problems of reduced contrast, blurred details, and color distortion in underwater images caused by the scattering and absorption of light via water molecules and suspended particles, an underwater image enhancement algorithm based on the Cycle-Consistent Generative Adversarial Network (CycleGAN) with fused skip features is proposed in this paper. The algorithm selects the first three layers of features from the down-sampling network, with sizes of 128, 64, and 32, respectively, fuses these features, and adjusts the feature maps adaptively by combining the channel attention mechanism and spatial attention mechanism. This adjustment enhances important features and reduces the interference of redundant features. Subsequently, these processed fused features, together with the original-scale feature maps, are used as skip connection features for the decoding operation of the decoder. Experimental results show that the images generated by the proposed method have higher sharpness and more natural colors, while effectively reducing the blue-green tone influence caused by water. Compared with the original images, the enhanced images using this method achieve improvements of 22.3%, 30.0%, and 17.3% in the unsupervised image quality evaluation metrics UIQM, UCIQE, and NIQE, respectively. In comparison with the underwater image enhancement algorithm based on the CycleGAN, the proposed algorithm achieves improvements of 0.34%, 13.7%, and 11.4%, respectively in the aforementioned three metrics. Through the effective integration of multi-scale information, the feature utilization efficiency is significantly improved, thereby verifying the important role of fused skip features in the field of image enhancement.