| Nowadays,people can acquire various different types of images more easily with the rapid development of information technology.Image is widely used in industrial production and daily life as a major carrier for information.A full-scope clear image can provide more abundant information for the task of human visual perception and image analysis.Nevertheless,in real shooting situations,only the targets within the depth of field can be clearly imaged due to the optical depth of field limitation of imaging equipment,while targets outside the scope often appear blurred.Multi-focus image fusion technology can integrate images captured within different depth-of-field ranges in the same scene to generate an image containing all targets in the scene clearly.The central problem of what this technique does is how to preserve the edge and detail information from the source image in the fusion result.In this article,further research on the multi-focus image fusion problem is launched in the context of deep learning techniques,and two new algorithms are proposed aiming to improve the performance of fused images.The main research work includes:The decision maps of conventional approaches are available by using the focus detection algorithm on single-channel gray-scale images.Motivated by the multi-scale idea,the traditional sun of Laplacian energy focus detection algorithm is improved in this thesis so that it can perform focus detection on multi-channel depth features and aggregate to create the final decision map.The method first trains an encoder-decoder network to extract the deep features of the source image.Then the improved sun of the Laplacian energy algorithm is used to perform focused detection on the deep features to obtain the decision map.Finally,we obtain the fused result based on the decision map.The proposed method nicely combines the advantages of traditional methods and deep learning,and it analyzes the depth features of the source image instead of the features of the source image extracted based on the traditional manual design.Experimental results demonstrate that the proposed method achieves state-of-the-art fusion performance in terms of both objective and subjective evaluation over existing methods.Furthermore,most existing spatial domain-based and deep learning fusion algorithms often need to generate an intermediate decision map in the fusion process,and a series of post-processing steps are required to refine the decision map to obtain a good fusion result.The fusion results produced by these algorithms done with several steps are not only easy to lose the information of the input source image details and make the boundary regions distorted,but also the algorithms do not operate efficiently.To address this problem,an attention mechanism-based multi-focus image fusion model is proposed in this thesis in order to improve the fusion efficiency of the fusion framework and obtain high-quality fusion results.At first,it is considered that currently available datasets are generated from image slices and do not contain real fused image labels.Therefore,a synthetic method containing the generation of true fused image labels is introduced,following which a large-scale Dataset is generated to train the proposed model.The overall procedure of the model follows:the features of the original image are first extracted using multiple convolutional layers,then the features are fused in the channel dimension and spatial dimension using the designed fusion strategy,and finally the final fusion results are generated by reconstructing the fused features.But on the other hand,since existing deep learning-based multi-focus fusion algorithms perform feature fusion using simple fusion strategies for processing image features,which ignores the fact that the importance of different source image features in the fused image is different,resulting in difficulties in restoring a high-quality fused image.Hence,an attention-based fusion strategy is proposed,and it improves the flexibility when fusing features by assigning different importance to different features in the channel dimension and the spatial dimension.This model also has three sections,but all three sections are consecutive and without steps such as decision maps and postprocessing.It is a typical end-to-end network model that effectively improves the quality of fusion results and optimizes the fusion efficiency of the network.Experimental results show that the proposed algorithm is excellent and efficient,taking only 0.26 seconds to fuse each pair of source images. |
