Underwater image enhancement based on lightweight non-linear activation free network
To overcome the common issues of color distortion and low contrast in underwater image processing,this study developed an innovative underwater image enhancement technique based on a lightweight non-linear activation free network.The core feature of this technique is the use of multiple cascaded non-linear activation free modules without traditional activation functions,significantly enhancing the flow of information and the efficiency of feature extraction.Additionally,the model integrates an innovative layer attention mechanism,which effectively identifies and optimizes feature dependencies between different layers,enhancing the expression of key information through dynamic adjustment of feature weights.To comprehensively evaluate the performance of the proposed method,detailed experiments were conducted on the Large Scale Underwater Image(LSUI)dataset.Compared with leading models such as FUnIE-GAN and Shallow-UWnet,our model demonstrated a significant advantage in structural similarity index(SSIM),with improvements of 8.17％and 4.13％respectively,markedly enhancing the color accuracy and detail retention of the images.Furthermore,the parameter count of our model was significantly reduced,decreasing by 98％and 50％respectively,greatly enhancing the model's practicality and deployment capabilities in environments with limited computational resources.The results of this study confirm the effectiveness of this enhancement technique in addressing key visual challenges in underwater imaging and also demonstrate its potential for application in extreme visual environments.By introducing this lightweight and efficient image enhancement approach,new pathways have been opened for the further development and innovation of underwater image processing technologies,laying a solid foundation for the widespread deployment of underwater vision systems in practical applications.
万方数据
维普
