频域与空间域协同的生成对抗网络图像去雾算法OA

Foggy weather degrades the quality of images acquired by imaging devices,further increasing the difficulty of advanced vision tasks.Current defogging methods rely on foggy and non-foggy image pairs,and acquiring near-perfect data is difficult.The learning results based on this data model are poor adaptability and robustness in the real world.Based on domain migration,the frequency and spatial domain synergistic cycles-consistent generative adversarial network image defogging algorithm(DWTF-CycleGAN)is proposed.A discrete wavelet transform-based Transformer module DWT-Former is constructed in the generator,containing high-frequency branches and low-frequency branches.The high-frequency branch assigns higher weights to texture and edge information through a frequency-domain guided multi-head self-attention mechanism.On the contrary,the low-frequency branch contains low-frequency subdomains and upper image features.The output of the high-frequency branch is connected to the upsampling module via a jumper,and the out-put of the low-frequency branch goes directly to the next segment of the encoding process.A method using adaptive atmo-spheric light estimation is proposed to compute image haze density for evaluating haze distribution.Furthermore,a cyclic haze density loss function is introduced to constrain the generator,guiding the network to focus on dehazing as the optimi-zation objective and to reduce deviations during the domain transfer process.Experimental results show that proposed fog removal method improves the average gradient compared to the current mainstream fog removal algorithms.The FADE and NIQE values decrease significantly,the image details become rich,and the defogging is improved.