| Infrared and visible image fusion technology can effectively fuse the features of two kinds of images to overcome the limitations of a single image in different environments.Through the effective combination of the two kinds of image feature information,the target details can be sufficient and the scene can be clear.In the case of bad weather,the required target can be imaged in real time and accurately.At present,infrared and visible light fusion has been widely used in military,security,space,medicine,target recognition and other fields.Therefore,this paper studies and improves the algorithm for the application of pulsed coupled neural network and convolutional neural network in the two kinds of image fusion,so as to improve the effect of final image fusion and ensure the applicability and timeliness of the algorithm.The main work is as follows:Aiming at the problems of incomplete contour information,missing edge and texture details in infrared and visible image fusion caused by using the same fusion strategy for high and low frequency subbands in traditional algorithms,An improved Simplified Pulse Coupled Neural Network（MSPCNN）and Fuzzy C-Mean（FCM）image fusion algorithm is proposed.First,infrared and visible images were decomposed into high and low frequency sub-bands image using the Non-Subsampled Shearlet Transform（NSST）.Then MSPCNN is used to fuse the decomposed high frequency subband,and a Gaussian distribution weight matrix is used for processing to enhance the detail information and contrast.Then,the obtained low-frequency subband images were extracted by using FCM clustering algorithm,and the approximate threshold of clustering center was set to simplify the process to achieve low-frequency subband fusion.Finally,NSST is used to achieve infrared and visible image fusion.Through statistical calculation of simulation experiment results,the objective reference indexes of AVG,SSIM,QAB/F,PSN,SF and FD in the proposed method increased by 32.35%,5.34%,6.46%,17.09%and 7.97%on average,which were consistent with subjective visual observation.Due to the improved running speed and timeliness of the simplified algorithm of model parameters,the algorithm is more suitable for complex scenarios.Aiming at the problem that traditional convolution can extract single and insufficient features from infrared and visible images,this paper proposes a fusion algorithm of infrared and visible images based on the combination of empty convolution and attention mechanism.Considering the different feature properties of infrared and visible images,a three-layer network is constructed by shallow feature extraction,deep feature fusion and fusion sub-network respectively processing the data with large differences to ensure the full extraction of features.During the training,the images of two modes were input at the same time,and the hollow convolution structure was used to extract the initial feature information.The shallow feature information was fully extracted under different receptive field sizes.Secondly,the three-layer dense convolution structure is used for deep feature fusion to preserve as much as possible important feature information in infrared and visible images and Concat is used to complete the feature fusion.Finally,CBAM attention mechanism is used to focus the features of infrared and visible images from both channel and space,so as to ensure that all significant feature information can be used in the fusion,improve the performance of network fusion and obtain the final fusion image with good effect.According to the statistical calculation of objective reference indicators,the algorithm improved 9.53%,45.63%,20.47%,36.43%,39.12%,52.14%on the EN,AG,MI,SF,SNR and CC objective evaluation indicators on average,which was consistent with the subjective visual evaluation,proving the effectiveness of the algorithm. |
