Infrared image can capture face information more clearly in the case of poor light conditions.It is often found that cross do-main recognition occurs in practical scenarios,this paper proposed an improved cross domain of visible to near-infrared face image conversion technology based on antagonism generation network.The generator component of the cyclical adversarial generative network has been refined by adding a facial classification branch and incorporating facial discrimination attribute constraints,enabling the gen-erated images to effectively preserve facial identification details.Additionally,the integration of a non-local algorithm allows the model to automatically focus on key facial features,enhancing the quality of cross-domain face generation.To ensure effective mapping during training,a cyclic consistency loss is proposed,particularly for larger networks.Experimental results on public datasets demonstrate im-proved performance of the proposed approach.
维普
万方数据