基于生成对抗网络的红外与可见光视频语义特征驱动融合方法OA

To address the performance bottleneck in existing video fusion methods—which often priori-tize visual quality improvement while neglecting the requirements of downstream detection tasks and fail to achieve differentiated processing of target regions,a semantic feature-driven fusion method for infrared and visible light videos based on GAN is proposed.A semantic feature embedding module(SFEM)is de-signed to effectively integrate semantic features into the fusion network.This module calculates the simi-larity between source and target images in the feature representation space by using dynamically generated semantic-guided convolutional kernels and attention mechanisms.A dual-path feature extraction archi-tecture is constructed,comprising a texture feature extraction module(TFEM)and a contrast feature ex-traction module(CFEM).These modules separately extract contrast and detail information from infrared and visible light images,achieving deep fusion of texture,contrast,and semantic features,thereby enhan-cing downstream task performance.During the training phase,a target detection sub-module is integrat-ed.The detection loss is backpropagated into the fusion network,enabling task-oriented optimization of the fusion process.Experimental results demonstrate that this method outperforms other comparative ap-proaches in balancing pixel intensity and preserving target object texture,while also exhibiting significant advantages in downstream target detection tasks.