Research On Multi-Focus Image Fusion Based On Multi-Scale Auto-Encoder Network

As modern information technology develops increasingly,multi-source images have been gradually applied in our daily life.By the means of detecting and fusing effective complementary features in multi-source images,more accurate and comprehensive information expression can be achieved by the image fusion technology than the single sensor.The fused image is more true of human vision perception,so that it also contributes to the subsequent task processing such as human observation and computer vision applications.In the sub-field of image fusion,Multi-focus image fusion technology is a popular research focus.Its application prospects are promising and broad in digital photography,bio-medicine,machine vision and other fields by effective and low-budget expanding the depth of the opticallens.On the basis of the previous research outcomes of multi-focus image fusion task,this paper focuses on further analysis and investigation.With the goal of the problems involving insufficient feature extraction of the current fusion algorithm and the dropout of detailed texture information,combined with the most advanced learning performance in deep learning,the traditional theory and technology have been expanded and improved accordingly.That is to say,it combines the advantages of the powerful feature extraction ability of the self-encoder network and multi-scale transformation in extracting and combining image features on different scales,and proposes two multi-focus image fusion algorithms.1.Aiming at the problems of single feature extraction scale and insufficient feature utilization in the current fusion method based on deep learning,a multi-focus image fusion method is proposed on account of multi-scale convolution self-encoder network,which combines the advantages of CNN’s powerful feature extraction ability and DWT combined image features on different scales.Firstly,in order to integrate depth features and fine-grained features,the encoder stage introduces Dense connections and Sobel operator to promote the learning of image gradient features.Secondly,the fusion layer brings in a multi-scale system,uses the DWT to transform the feature map into a wavelet domain,and meanwhile,adaptive fusion rules for low-frequency and highfrequency components are designed by means of the attention mechanism,as a result,the fusion performance gets promoted.Seven representative multi-focus fusion methods are selected for comparison,and the ability of the algorithm in image feature extraction is confirmed by qualitative and quantitative analysis.2.With the purpose of further optimizing the fusion performance and perfecting the quality of the fusion image,a multi-focus image fusion method on the basis of transformation self-supervised network is presented.Firstly,Some zones of the input image are transformed and reconstructed in the training stage.In the transformation domain,the low frequency and high frequency are treated with Gaussian blur and Rigid transformation respectively,so that the network will pay more attention to multi-focus image semantic information and edge detail information tasks during the reconstruction training process.Secondly,bring in Transformer module and combine it with Sobel module in order to enhance the ability of network feature extraction,which covers the shortage of establishing remote correlation in the existing CNN-based method with the superior capacity of Transformer to capture remote relationship by means of selfattention mechanism.Experiments have proved that this method can make full use of local and global information,have higher contrast of fusion images,stronger sense of spatial hierarchy,and excellent visual effects.It has been confirmed that the performance of the fusion algorithm in this paper meets the expectations,and it can take advantage of local and global information,and the visual effect is excellent.