Research On Multi-focus Color Image Fusion Algorithm

Multi-focus image fusion is a digital image technology that fuses multi-focus images in the same scene into an image.This technology aims to solve the problem of inability to full focus due to the limited focal length of optical lenses.The algorithmic steps in traditional image fusion algorithms are feature extraction,feature comparison,and implementation of a fixed fusion strategy.The advantage of deep learning is that it can automatically learn the inherent laws of data samples and reduce artificial interference of factors.The current deep learning scheme applied in multi-focus images regards this problem as a classification problem,which requires the help of digital image processing techniques as post-processes to assist image fusion.However,this paper treats the problem as an image-generating regression problem,focusing on a fully convolutional deep learning network that directly fuses multi-focused images without the use of subsequent post-processing.The work in this paper covers the following aspects.(1)this paper proposes a multi-focus image fusion method based on auto encoder.The multi-branch structure and dense network components are used to extract shallow features and deep features from multi-focus images,and then fuse the features and reconstruct image.Experimental tests verify the feasibility of the idea and the effectiveness of the fusion algorithm.At the same time,the algorithm can well extract and retain details such as textures and edges in multi-focus images.Compared with other algorithms,the fused images in this paper have high definition and the boundary of fusion is natural.(2)this paper proposes a multi-focus image fusion algorithm based on generative adversarial networks.The generator model and the discriminator model are used to construct the deep learning network,and the ECA module is embedded in the generator model.In the experiments,three datasets such as multi-focus image pairs,random mask image pairs and parallel mask image pairs are constructed for network training.The experiments show that the masked image pairs can tap the network performance better and outperform the other compared algorithms in the objective index evaluation.At the same time,a control experiment of embedding the attention mechanism is carried out,which shows that the ECA module can improve the performance of the network well,and the effect is better than that of the embedding SE module.In the fused image,the full-focused images fused by our algorithm have the characteristics of fine texture,high fidelity,clearness and nature.