Research Of Multi-focus Image Fusion Algorithm Based On Convolutional Neural Networks

Multi-focus image fusion(MFIF)technology is widely used in the digital photography,medical imaging,and security monitoring.The task of MFIF aims to integrate multiple images with different focus areas into one all-in-focus image.The MFIF algorithms based on convolutional neural network(CNN)have achieved great improvement.However,there are still some problems in existing methods,such as multi-scale features extraction of source images,and application of features at different layers of CNN.To overcome the above-mentioned problems,this paper studies on the deep fusion network in terms of deep features extraction of source images and network architecture.These researches can improve the performance of MFIF and have great importance in practical applications.The main studies are presented as follows:(1)A multi-focus image fusion algorithm based on multi-scale dilated convolutional network is proposed.According to the problem of single-scale feature extraction from source images,this paper proposes a multi-scale dilated convolution module,which uses dilated convolution with different dilation rates to extract multi-scale features.Such operation can improve the performance of multiscale feature extraction.Furthermore,the SE(Squeeze-and-Excitation)attention module is used to make the image feature extraction and analysis more accurate and improve the quality of the fused images.The experimental results show that the proposed algorithm can effectively extract the multiscale features from the source images,improve the representation of network,and provide better fusion results than the comparised algorithms.(2)A multi-focus image fusion algorithm based on dense attention network is proposed.According to the loss of some useful information in the middle layer of the network,this paper designs a dense convolutional module.The developed dense CNN can enhance the propagation of features and reuse the features at middle layers.In addition,the SE attention module is employed to highlight the importance of salient features.The loss function contains the perceptual loss,similarity loss and mean square error loss.The experimental results show that the proposed method can preserve the details information effectively.(3)A multi-focus image fusion algorithm based on multi-scale dilated residual network is proposed.To completely use the multi-level salient features extracted by the network,this paper propose a multi-scale dilated residual module,which consists of the residual module and the CBAM(Convolutional Block Attention Module)module to extract the multi-scale feature.Then,the skip connection is adopted to combine the salient information at multi-level of network,which can improve the quality of fused image.The experiment comparisons on various images show that the proposed algorithm can effectively use the salient information at multi-level of network,and achieves better results in terms of detail retention,contrast and sharpness.