Improvement Of Panchromatic Remote Sensing Image Processing Algorithms And Optimization Of Target Detection And Recognition

Panchromatic remote sensing image has a large amount of information,rich edge details,and contains important perceptual information.In order to extract the region of interest in the image and detect and recognize the target more accurately and effectively according to the characteristics of the target,this paper studies the improved image edge detection,segmentation and target detection and recognition algorithms,respectively,in order to meet the needs of panchromatic remote sensing image target detection and recognition.The aim of improving the quality of edge information extraction is to automatically determine the number of categories and segment the image globally,optimize the image detection and recognition algorithm,and improve the accuracy of target recognition.Firstly,because of the problems of over-detection,missing detection,error detection and weak anti-noise ability in texture edge detection of remote sensing images using gradient variation,an edge detection algorithm combining fractional differential difference and Gauss curvature filtering is proposed.The gradient field of panchromatic remote sensing images is enhanced by fractional differential difference operation,and the non-linear image expansion is smoothed by Gauss curvature filtering.In the scattered part,the steepest descent point of regularization energy is found to optimize the fractional order and iteration times in order(0-2)in the differential process,so as to improve the quality of edge information extraction of noisy images.The experimental results of remote sensing images show that the algorithm can suppress background pseudo-noise caused by non-linear amplification and diffusion of noise in the process of texture edge extraction of remote sensing images,and retain rich image texture edges.Information.Secondly,in panchromatic remote sensing image segmentation,an image segmentation algorithm is constructed by combining RJMCMC and SA(Reversible Jump Markov Chain Monte Carlo and Simulated annealing,RJMCMC+SA).Under the condition of ensuring the complexity and accuracy of remote sensing image segmentation model,the number of segmentation categories can be automatically determined;the image is processed by Gauss Curvature filtering(GC).The image is smoothed geometrically and the parameters in the nonlinear regression model are formalized according to Bayesian theory to establish the posterior probability distribution.Then the posterior probability distribution is realized by RJMCMCMC+SA algorithm,and the number and parameters of the radial basis function are determined.The automatic determination of the number of categories and the global image segmentation are completed.For panchromatic remote sensing images,the posterior probability distribution is segmented with four radial basis functions,respectively.The model is compared with four segmentation algorithms.The results show that the algorithm not only achieves a good balance in complexity and accuracy,but also can automatically determine the number of image categories.Finally,the idea of completing the target detection and recognition process is to use the fractional differential difference and Gaussian curvature filter edge detection algorithm proposed in this paper to select the aircraft segmented by RJMCMC+SA algorithm and make the neural network training data set.Based on the Faster R-CNN network,a cyclic attentional convolutional neural network(RA-CNN)was introduced.Using the RPN network and the APN network,the attention frame and the multi-scale region were recursively learned by using the attention mechanism to generate candidate frames and mutually reinforcing methods.Discrimination and feature representation.RPN shares the convolution feature with Fast R-CNN,and the RPN network uses the anchor frame mechanism to generate candidate regions.After ROIPooling processing,the feature map of 7*7 size is obtained,which is sent to the fully connected layer for classification and regression.The Faster R-CNN The target obtained by the network processing is sent to the RA-CNN,and the multi-scale network shares the same network architecture.The learning of each scale network including the classification subnet and the APN network can ensure sufficient recognition capability,and the softmax loss and attention through the classification network.A confidence score for pairwise sorting loss between network sizes is proposed to optimize the loop network.Compared with the Faster R-CNN network,the improved model based on Faster R-CNN and RA-CNN not only achieves image target detection,but also improves target recognition accuracy by 18.6%.