Research On Supervised Deep Learning Image Fusion Algorithm Based On Attention Mechanism

Image fusion is the fusion of salient information from different sensors in the same scene to produce a high quality fused image containing all the important features.The fused image provides a multifaceted description of the scene for subsequent processing tasks and is now widely used in robot vision,satellite imaging and other applications.Multi-focus image fusion is a technique for fusing images containing different focus regions to obtain a fully focused image.In the fusion of infrared and visible images,the infrared image can overcome the environment such as smoke and occlusion to identify the thermal target information,and the visible image contains rich texture information,and the two can be combined into one to obtain a composite image containing more features.In this paper,a supervised deep learning image fusion method based on an attention mechanism is proposed,which can adaptively refine the intermediate feature mapping,strengthen important features and weaken unimportant features by embedding an attention mechanism in the deep learning model,thus significantly improving the quality of the fused image.Extensive objective and subjective experimental comparisons show that the method in this paper has better performance.The main work of this paper includes.(1)A supervised convolutional neural network model based on an attention mechanism is proposed.The model avoids the problem of gradient disappearance while deepening the network to extract deep features through 12 Res Net residual blocks.The obtained intermediate layer features are further enhanced with useful features and suppressed with useless features in both channel and space dimensions through the attention mechanism blocks.The supervised training of the network with the dataset generated in this paper has resulted in a high quality improvement for fused images.(2)The proposed network model is improved to suit different fusion tasks.For the multi-focus image fusion task,a two-channel network is used.Two source images are connected on the channels and the complementary source images are used as input for pairwise learning to obtain the focus map of the source images.The focus map is multiplied by the source image to obtain the final fused image.In the IR and visible image fusion task,a pseudo Siamese network is used to extract features from the IR and visible images respectively,and the final fused image is reconstructed.Appropriate networks were used for different fusion tasks,which significantly improved the quality of the fused images.(3)Extensive experimental results show that the method in this paper has better performance in the fusion of infrared and visible images and multi-focus image fusion.The subjective and objective comparison results significantly outperform popular deep learning algorithms.