Lightweight image fusion algorithm based on generative adversarial networks
To address the issues of high algorithm complexity,large storage space requirement,heavy computational resource consumption,and deployment difficulty associated with conventional image fusion algorithms due to their big size,a lightweight image fusion algorithm was proposed through the integration of depthwise separable convolution(DSConv)and the convolutional block attention module(CBAM)in the target-aware dual adversarial learning(TarDAL)network.The traditional convolution layers in the generator structure were decomposed into depthwise convolution and pointwise convolution,forming DSConv to achieve a lightweight algorithm and reduce computational cost and the number of parameters.The features extracted from the source images were processed by the CBAM mixed attention mechanism,which allowed the network to focus more on important channel and spatial features during feature learning,enhancing the expression ability of the fused image.Test results on the M3 FD dataset indicate that,compared with the original TarDAL algorithm,the lightweight algorithm better preserves details and texture in subjective evaluation,while reducing interference from strong light.For the objective evaluation,changes in image evaluation indicators are minor.The lightweight model can generate fused images with good quality,with a parameters reduction of 86.24％compared with the original TarDAL algorithm.
万方数据
维普
