Search

Denoising is an important preprocessing step for the effective application of remote sensing images in mining area. The existing remote sensing image denoising methods based on statistics, domain transformation and learning generally have the problems of excessive smoothing of details and insufficient texture preservation. Based on the good edge-preserving property of guided filtering, an iterative guided filtering method is proposed. The method enhances the edge characteristics extraction effect of remote sensing images by guided mapping of residual information, and iteratively performing guided filtering and hyper-parameter shrinkage. The iterative guided filtering is combined with traditional wavelet soft threshold, non-local mean(NLM) filtering, block matching 3D(BM3D) filtering and other denoising methods, which improves the peak signal-to-noise ratio of the traditional method effectively. Among them, NLM and BM3D filtering have the most obvious effects on improving the denoising performance. The iterative guided filtering and BM3D filtering are fused, and the denoised images are initially obtained through BM3D filtering to obtain residual data. The iterative guided filtering is used to process the residual data. While improving the image denoising effect, the image detail characteristics are well preserved. The iterative guided filtering and BM3D filtering fusion method are used for coal gangue yard identification and landslide area edge recognition in remote sensing images of mining areas, and good results have been achieved.

Tongue segmentation is an essential part of intelligent medicine diagnosis. The purpose is to generate an accurate contour of the tongue region by a precise mask. In recent years, deep learning methods have been widely applied in the field of image processing and have achieved impressive performance. With the increasing requirement of the performance for medical image segmentation, many scholars have employed deep learning to tongue segmentation. The methods of deep learning-based tongue segmentation are analyzed, classified and summarized. In the field of tongue segmentation applications, various tongue segmentation methods based on deep learning are divided into eight types: convolutional neural network (CNN), fully convolutional network (FCN), convolutional model with graphical model, encoder-decoder based model, regional convolutional network-based model, atrous convolutional model, transfer learning and other methods. This paper presents a comprehensive survey of the recently developed deep learning for tongue segmentation, and analyzes the strengths and weaknesses of these methods. The commonly used data sets and evaluation indexes of tongue segmentation based on deep learning are visually compared and quantitatively evaluated. This survey is concluded by discussing the potential development in future research work.

To solve the problem of matching difficulty caused by the significant nonlinear grayscale differences between SAR and optical images, this paper proposes a fast matching algorithm based on image structural properties named SOFM(SAR-to-optical fast matching algorithm).The traditional methods based on image grayscale are generally difficult to resist the nonlinear grayscale differences between SAR and optical images, but the geometric constructs and shape features can exist stably among different types of images, so in our the proposed method both the magnitude and orientation information of image gradient are used to build a geometric structural feature descriptor named AWOG(angular weighted orientated gradients), then based on the template matching strategy, the sum of squared difference of the descriptors is used to define the similarity metric for matching and then the image matching function expressed in the frequency domain is given. A complete set of image matching process is established based on SOFM, and has been validated using multiple pairs of SAR and optical images, the results show that the proposed method can effectively resist the nonlinear grayscale differences between SAR and optical images, and outperforms the traditional classical image grayscale-based methods and existing image structural-based methods in matching performance and precision.

In order to solve the problem that the parameters of pulse-coupled neural network can't be adjusted adaptively in pan-sharpening image fusion, a non-subsampled shearlet transform remote sensing image fusion method based on the combination of parametric-adaptive pulse coupled neural network model and energy-attributing fusion strategy is proposed. First, the high and low frequency coefficients are obtained by extracting the Y luminance component of the multispectral image YUV color space transform and transforming it with the panchromatic image. Then, aiming at the low-frequency sub-band coefficients are fused by the EA method, the high-frequency sub-band coefficients are obtained by the PA-PCNN model to determine the optimal PCNN model, and then the high-frequency sub-band coefficients are fused; finally, the fusion image is obtained by inverse transformation of NSST and YUV. In this paper, six objective quality indexes, such as spatial frequency, relative dimensionless global error, ERGAS, correlation coefficient, visual information fidelity for fusion, gradient-based fusion performance and structural similarity index, are selected to evaluate the spectral and spatial detail information of the fused images, compared with SE, DGIF, COF and PA-PCNN fusion methods, the proposed method is validated by using multiple sets of high-and low-resolution panchromatic and multispectral remote sensing images, the results show that this method is generally superior to the traditional fusion method of panchromatic and multispectral remote sensing images in objective evaluation and visual perception.

In order to solve the problems of small field of view and unclear detail features of monitoring images in coal mines, a multi-view image stitching method in mine is proposed. Firstly, an improved adaptive histogram equalization method with contrast limitation is used to pre-process the images to highlight the image details and improve the contrast. Secondly, the ORB algorithm is used to extract the image feature points, and the improved Brief algorithm is used to calculate the feature descriptors. Thirdly, the K-nearest neighbor(KNN) algorithm is used to achieve rough matching of feature point pairs, and the random sample consensus(RANSAC) algorithm is used to filter and eliminate the mismatched feature point pairs, and the optimal perspective transformation matrix is solved to transform the coordinates of the pixel points of the image to be matched. Finally, the hat function weighted average fusion algorithm is used to stitch and fuse the fixed image and the image to be matched. The experimental results show that compared with the speeded up robust features(SIFT) and KAZE algorithms, the ORB algorithm reduces the number of feature points extracted for a single image by 48% and 33% respectively, which improves the effective feature point extraction capability. The feature point extraction time is reduced by 17% and 34% respectively, which improves the calculation efficiency. The images stitched by this method avoid the phenomenon of cracks and black lines at the joints, the image transition is natural, and the definition is high.

In order to solve the problem of inaccurate feature point extraction and poor matching effect of existing underground image matching algorithms, an underground image matching algorithm combining homomorphic filtering and histogram equalization is proposed. The image is sharpened by homomorphic filtering to improve the image clarity, and the image is processed by the contrast-limited adaptive histogram equalization (CLAHE) algorithm to highlight the edge detail information of the image and improve the image contrast. In order to solve the problem of mis-matching in the traditional AKAZE algorithm, on the basis of rough matching by the brute force matching algorithm, the random sampling consensus (RANSAC) algorithm based on the homography matrix is used to perform accurate matching and eliminate the mis-matched point pairs. The experimental results show that using single-parameter homomorphic filtering and CLAHE algorithm to enhance the image can stretch the gray level of the image, reduce the number of dark pixels and increase the number of bright pixels, which makes the gray level distribution smoother and helps to preserve the details and boundary information of the image. Using RANSAC algorithm based on the homography matrix for accurate matching can detect more feature points and improve the matching accuracy. The matching effect is better than that of SURF algorithm and traditional AKAZE algorithm with a maximum of 96.09%.

The wavelet transform cannot maintain a balance in image edge preservation and detail processing and the multi-scale Retinex algorithm is prone to halo artifacts and serious noise pollution in images. In order to solve the above problems, a mine low illumination image enhancement algorithm is proposed by combining the wavelet transform with the multi-scale Retinex algorithm based on multi-scale guided filtering. Firstly, the algorithm decomposes the low illumination image into wavelets to obtain the high-frequency and low-frequency components. Secondly, the wavelet denoising is applied to the high-frequency components of the image using a three-stage threshold function, and nonlinear global luminance correction is applied to the low-frequency components of the image to enhance the image luminance. Moreover, the multi-scale guided filter function is used to estimate the illumination components instead of the Gaussian filter function of the traditional multi-scale Retinex algorithm, and then the reflection components are obtained. The principal component analysis method is used to fuse the reflection component and the non-linear global luminance correction image so as to improve image edge detail preservation effect effectively. Finally, the wavelet reconstruction is performed on the high-frequency components and low-frequency components of the image, and the wavelet reconstructed image is nonlinearly transformed to solve the image graying problem. The experimental results show that the algorithm has strong noise suppression capability, can improve the image luminance and contrast effectively, make the image edge preservation performance and detail information richness effectively balanced, and avoid the image halo artifacts and color distortion.

Existing based on convolutional neural network classification method of hyperspectral images usually rules of the square area of image convolution, not widely adapt to different terrain local area distribution and geometry appearance of the image, therefore, under the condition of small sample classification performance is poorer, and figure convolution can network topology information on the map represent irregular image area of the convolution. Therefore, a hyperspectral image classification method based on graph convolution network is proposed. In this method, the spatial spectral information of the image is considered in the process of constructing the graph, and the feature information of the neighbor node is aggregated by the graph convolution network. Experimental results on three data sets, Pavia university, Indian Pines and Salinas, show that this method can achieve a high classification accuracy with a small number of training samples.

Due to the high cost and time-consumption of artificial semantic tags, domain-based adaptive semantics segmentation is very necessary. For scenes with large gaps or pixels, it is easy to limit model training and reduce the accuracy of semantic segmentation. In this paper, a domain adaptive semantic segmentation network (DA-SSN) using the improved transformation network is proposed by eliminating the interference of large gap pictures and pixels through staged training and interpretable masks. First, in view of the problem of large domain gaps from some source graphs to target graphs and the difficulty in network model training, the training loss threshold is used to divide the source graph dataset with large gaps, and a phased transformation network training strategy is proposed. Based on the ensurance of the semantic alignment of small gap source images, the transformation quality of large gap source images is improved. In addition, in order to further reduce the gap between some pixels in the source image and the target image area, an interpretable mask is proposed. By predicting the gap between each pixel in the source image domain and the target image domain, the confidence is reduced, and the training loss of the corresponding pixel is ignored to eliminate the influence of large gap pixels on the semantic alignment of other pixels, so that model training only focuses on the domain gap of high-confidence pixels. The results show that the proposed algorithm has a higher segmentation accuracy than the original domain adaptive semantic segmentation network. Compared with the results of other popular algorithms, the proposed method obtains a higher quality semantic alignment, which shows the advantages of the proposed method with high accuracy.

An effective algorithm for unmixing hyperspectral and panchromatic images of non-negative matrix factorization based on spectral reconstruction constraint is proposed.Firstly,this algorithm employs the regularization with minimum spectral reconstruction error in the process of non-negative matrix factorization for the hyperspectral image,and searches for the optimal regularization parameter through multi-objective optimization to inspire the spectral signature matrix to contain more real spectral features.Then,the panchromatic image is factorized by non-negative matrix to obtain the abundance matrix with the details of the image.Finally,the fusion result is reconstructed by using the spectral signature matrix and the abundance matrix.The experimental results show that the fusion result of the proposed algorithm maintains more details of panchromatic images and effectively decreases spectral distortion simultaneously.It has better performance in both visual effects and objective evaluation than traditional algorithms.

Due to the interference of instrumental noise,hyperspectral images (HSI) are often corrupted to some extent by Gaussian noise,which will seriously affect the subsequent performance of image processing.Therefore,image denoising has been considered as an important pre-processing step.Besides,due to the high dimensionality of hyperspectral data,the running efficiency is also a critical factor along with the visual evaluation.For the sake of improving both the efficiency and efficacy,we first project the high-dimensional hyperspectral image into certain spectral subspace,and then learn an orthogonal basis matrix.On that basis,the spatial non-local similarity and the global spectral low rank property of hyperspectral are jointly introduced to denoise the low-dimensional subspaces.Finally,all the restored low-dimensional image can be used along with the orthogonal basis to recover the original HIS data.Among these steps,the non-local denoising process first forms certain amount of tensor cubes by the non-local similarity,and followed by several tensor groups using the block matching method.In general,these groups enjoy strong low-rank essense due to the explicit neighborhood similarity.For better revealing the low-rank property of each tensor group,we propose a weighted and truncated nuclear norm by taking both the advantages of weighted nuclear norm and truncated nuclear norm.Moreover,an improved optimization scheme based on the accelerated proximal gradient is presented for a fast solution.Extensive simulation results show that our denoising scheme outperforms state-of-the-art methods in objective metrics and better preserves visually salient structural features.

In order to solve the problem that the fusion result is limited due to the over dependence on the label image when using the deep learning methods in the multi band image fusion field,a multi-band image feature-level self-supervised fusion method based on multi-discriminator generation adversarial network is proposed.Firstly,this paper designs and builds a feedback dense network as a feature enhancement module to separately extract multi-band image features and perform feature enhancement.Se-condly,it merges and connects the multi-band image feature enhanced results and reconstructs the fused image through the designed feature fusion module.Finally,the preliminary fused result and the source images of each band are input into the discriminator network respectively.Through the classification task of multiple discriminators,the generator is continuously optimized so that the output of the generator retains the characteristics of multiple band images at the same time to achieve the purpose of image fusion.Experimental results show that,compared with the current representative fusion method,the proposed method has better clarity,information volume,more detailed information,and is more in line with human visual characteristics.

Aiming at the low matching accuracy of local stereo matching algorithm in weak texture or discontinuous disparity areas, a stereo matching algorithm combining multi-scale fusion of convolutional neural network (CNN) and feature pyramid structure (FPN) is proposed. The feature pyramid is applied on the basis of the convolutional neural network to realize the multi-scale feature extraction and fusion of the image, which improves the matching similarity of the image blocks. The guide graph filter is used to quickly and effectively complete the cost aggregation. The disparity selection stage adapts the improvement dynamic programming algorithm to obtain the initial disparity map. The initial disparity map is refined so as to obtain the final disparity map. The algorithm is trained and tested on the image provided by Middlebury data set, and the result shows that the disparity map obtained by the algorithm has good effect.

To solve the problems of noise sensitivity and initial clustering center randomness in the segmentation of brain MRI images by FCM algorithm,an improved FCM image segmentation algorithm based on Tamura texture feature is proposed.Firstly,the Tamura texture feature of the image is extracted,and it is linearly weighted with the gray feature to form a fusion feature.Then,the density of pixel is calculated by using fuzzy neighborhood relation,and the initial cluster center is selected by combining it with distance relation.Finally,the fusion feature is used as a feature constraint for updating membership and clustering center.In the experiment,FCM,D-FCM,WKFCM and the proposed method are used to segment the images in Brain Web MRI dataset,and their anti-noise performance,accuracy and operation efficiency are compared.Experimental results show that the proposed algorithm can obtain better initial clustering centers,has better robustness in processing noise and gray inhomogeneity images,and can segment brain MRI images quickly and effectively.

In real scenes, due to the imperfections of equipment and systems or the existence of low-light environments, the collected images are noisy. The images will also be affected by additional noise during the compression and transmission process, which will interfere with subsequent image segmentation and feature extraction processes. Traditional denoising methods use the non-local self-similarity (NLSS) characteristics of the image and the sparse representation in the transform domain, and the method based on block-matching and three-dimensional filtering (BM3D) shows a powerful image denoising performance. With the development of artificial intelligence, image denoising methods based on deep learning have achieved outstanding performance. But so far, there is almost no relevant research on the comprehensive comparison of image denoising methods. Aiming at the traditional image denoising methods and the image denoising methods based on deep neural networks that have emerged in recent years, this paper first introduces the basic framework of the classic traditional denoising and deep neural network denoising methods and classifies and summarizes the denoising methods. Then the existing denoising methods are analyzed and compared quantitatively and qualitatively on the public denoising data set. Finally, this paper points out some potential challenges and future research directions in the field of image denoising.

There are significant nonlinear intensity differences between multi-modal images. Moreover, the noise in these images will cause image degradation. Therefore, the automatic registration of multi-modal images is a challenging task. To address the two problems, this paper proposes a multi-modal image automatic registration method, which is divided into two stages: pre-registration and fine registration. In the pre-registration stage, an improved SIFT algorithm is used to roughly align multi-modal images. In the fine registration stage, the block Harris detector is first used to extract evenly distributed feature points on the pre-registered reference image. Then, the structure information in the multi-modal images is captured by the anisotropic structure tensor to construct a feature descriptor, which is robust to noise. Furthermore, a similarity criterion named TOMI (tensor orientation and mutual information) is proposed combining the tensor orientation parallelism and gradient mutual information. Finally, Multi-modal images (including Optical, LiDAR, SAR, and Map data) are used to evaluate the proposed algorithm. The experimental results show that the method proposed in this paper is robust to nonlinear intensity differences and noise, and the matching effect is superior.

In recent years, scene classification of high-resolution remote sensing images based on deep convolutional neural networks has become the focus of attention. Because of the existing deep convolution neural network is not robust to the geometric deformation of remote sensing scene image, we proposed a novel scene classification method for high-resolution remote sensing image, based on the deep transfer deformable convolutional neural networks (DTDCNN). Specifically, the depth features of remote sensing image are extracted by using the trained depth model on the large-scale natural scene dataset (ImageNet), then, the deformable convolution layer is introduced to learn the depth features which are robust to the geometric deformation of remote sensing scene.The results show that:the accuracy of DTDCNN on AID, UC-Merced and NWPU-RESISC45 datasets is improved by 4.25%, 1.9% and 4.83% after adding the deformable convolution, respectively. By the adaptive adjustment of the receptive field for different objects in the scene, DTDCNN enhances the ability of spatial sampling position, and, effectively improves the accuracy of remote sensing scene classification.

The available wide-baseline image matching algorithms have been prone to failure or only producing few matches, due to the complex affine deformation and perspective distortion. On this basis, we proposed a novel affine invariant feature matching algorithm for oblique stereo images based on multivariate network. In our method, we applied the Hessian algorithm to extract initial feature regions, then we constructed triplet network (TN) model, and obtained affine invariant feature regions through deep learning. To improve the matching performance of similar features, we proposed multilateral constraint loss function to train multi-branch descriptor network (MDN) model, and then generated deep learning descriptors with higher discrimination for image features. Afterwards, the conjugate features were produced by the matching metric of nearest/next distance ratio (NNDR), and eliminated possible mismatch points through random sampling consistency (RANSAC) algorithm. Finally, experiments on oblique stereo images acquired by unmanned aerial vehicle verified the effectiveness of the proposed approach.

Aiming at the large amount of clutter and noise interference(mainly direct waves on the surface) in the original image of ground penetrating radar,an adaptive bilateral filtering method based on wavelet transform and K-Means singular value decomposition was proposed in order to effectively extract the target signal.The original radar data was decomposed by wavelet,and then the decomposed data was sparsely coded by K-SVD transform and updated atoms.The updated sparse coefficients and dictionary were used to reconstruct small blocks,and the small blocks were subjected to inverse wavelet transform to reconstruct the image.The reconstructed image was processed by adaptive bilateral filtering to obtain the processed image,so as to realize the complementary advantages of wavelet transform,K-SVD algorithm and adaptive bilateral filtering method.Compared with other methods,the experimental results show that the proposed method has a good clutter suppression effect and better performance in peak signal-to-noise ratio and target image entropy.The research results have certain application value in the actual radar survey tasks.

Aiming at the problem that traditional convolutional neural network (CNN) continuously stacks convo-lutional layers and pooling layers in order to obtain high accuracy, resulting in complicated model structure, long training time, and single data processing method, a optimized layered convolutional sub-health recognition algorithm of improved capsule network is proposed. Firstly, the original vibration data are transformed by wavelet denoising and wavelet packet denoising to better retain the useful information in the original signal for sub-health recognition. Secondly, CNN adopts the idea of layered convolution, parallelizes three convolution kernels of different scales, and carries on multi-angle feature extraction. Finally, the features extracted by the convolution kernels are input into the improved capsule network with pruning strategy for sub-health recognition. The improved capsule network can not only guarantee the accuracy, but also accelerate the sub-health recognition time, thus the problems of too compli-cated CNN structure and poor recognition effect are solved. Experimental results show that the proposed algorithm has high recognition accuracy and less recognition time.

The suspicious noise points are taken as center to construct a certain size neighborhood in this paper. Morphological close filtering is performed by pepper noise neighborhood. Then, morphological open filtering is applied by neighborhood salt noise point. The gray values of suspected noise points are replaced by filtered gray value of neighborhood center point. During the transforming process, gray values of non-suspicious noise points remain unchanged. Experiments show that this scheme is practical and plausible. The filtered image has advantages of less mean square error and high peak signal-to-noise ratio. Compared with the median filtering algorithm, the morphological open-close filtering algorithm，and the adaptive filtering algorithm, this proposed improved algorithm provides a better filtering property.

Denoising fitting of airborne laser bathymetry data is a key step in extracting the bottom terrain. The algorithm effects of wavelet adaptive threshold denoising, empirical model denoising (EMD) and joint denoising are compared in this paper, and then multivariate Gaussian fitting is used to test the denoising effect. The optimal denoising algorithm and parameter selection are obtained by comparison, and it is realized that the high-precision extraction of seabed features. This study has shown that: when the sounding data is denoised by wavelet threshold, the fixed threshold wavelet denoising effect is superior to other denoising effects, and the denoising decomposition level is more than 6 layers, which tends to be stable. The average accuracy of the algorithm reaches 8.218 2 after the fifth-order Gaussian fitting of the denoising data. The algorithm has strong robustness, it can meet the technical requirements of blue-green laser practical application, and provides a reference for accurately extracting seafloor feature information.

Aiming at the problem of poor precision performance of traditional Bayesian model saliency detection algorithm, a Bayesian model saliency detection algorithm based on improved convex hull was proposed. Firstly, the foreground of the image was extracted by the manifold ranking algorithm, which was used as the prior probability in Bayesian model. Secondly, Harris corner detection algorithm based on color enhancement was used to detect the feature points of the image in three color spaces of RGB, HSV and CIELab; the convex hulls in RGB, HSV and CIELab spaces were constructed respectively; and the intersection of convex hulls were obtained. Thirdly, the saliency region map was calculated by Bayesian model according to the prior probability and the observed likelihood probability obtained by combining convex hulls and color histograms. Finally, the proposed algorithm was tested in two public data sets MSRA and ECSSD. The experimental results show that the proposed algorithm can suppress the background noise effectively and detect the salient areas completely. The test results of F-measure value in MSRA and ECSSD databases are 0.87 and 0.71 respectively, and the accuracy-recall rate curve is higher than that of traditional classical algorithms in complex image databases. The proposed algorithm improves the detection effect of the traditional classical algorithm and the accuracy of saliency map detection.

Aiming at the issue of incomplete trajectories in the 2D epipolar image of circular light field, this paper proposes a 3D reconstruction method by using 3D Hough transformation. This method computes 3D point clouds by computing the parameters of feature trajectories in 3D image volume. By analyzing the 3D distribution of circular light field trajectories, binary curves in image volume are extracted, and their local orientation are further estimated by the 3D structure tensor. The 3D Hough space generation and the parameter selection method are proposed to the 3D curves detection. The parameters of these curves are converted to 3D point clouds on each view and then merged to final 3D reconstruction. The ambiguity of Hough transformation solution on 2D epipolar image is overcome by the 3D analyzing method. The experiments are carried out on both synthetic and real datasets. The experiment results show that this method can improve the reconstruction performance compared with the state-of-the-art in circular light field.

Aiming at the problems of nonlinear gray difference, speckle noise and different imaging viewpoints in SAR and optical image registration, this paper presents a SAR and optical images registration using uniform distribution and structure description-based ASIFT. In the proposed algorithm, firstly, the guided scale space is constructed by guided filter to achieve noise suppression and edge preservation. In the feature extraction stage, the phase congruency is utilized due to the nonlinear gray difference, and combined with scale space gridding to extract from images for the uniform feature points. In the feature description stage, the consistency gradient magnitude and orientation of SAR and optical image are calculated by extended phase congruency method, which improves the accuracy of the main orientation and descriptor. At last, Optimal-RANSAC is used to establish feature descriptor matching to achieve effective registration. The simulation experiment and analysis on four pairs of real images show that the proposed algorithm has more accurate registration accuracy than SAR-SIFT and traditional ASIFT.

Aiming at the problem that existing image saliency detection algorithms can't correctly detect salient objects in complex environments, this paper proposes a method combining sparse reconstruction error and the compactness of image salient regions to calculate image saliency. Firstly, the main structure of the image is extracted to reduce background noise. Then the processed image is segmented into several superpixels. On one hand, the background dictionary is constructed by using the boundary superpixels. Each superpixel is projected on the dictionary for sparse reconstruction, and the reconstruction error is used to obtain a saliency map based on sparse reconstruction. On the other hand, the maps based on foreground seeds and background seeds are calculated separately via compactness of salient objects and fused. Finally, the saliency map obtained by the sparse reconstruction error and compactness are fused to get the final saliency map. The proposed algorithm is compared with 13 algorithms proposed in recent years on several public datasets. The experimental results show that the proposed algorithm is superior to all comparison algorithms.

The method of using unmanned aerial vehicle (UAV) to obtain aerial image information of target scene has the characteristics of wide coverage, strong mobility and high efficiency, which is widely used in urban traffic monitoring, vehicle detection, oil pipeline inspection, regional survey and other aspects. Aiming at the difficulties of the object to be detected in the process of aerial image object detection, such as multiple orientations, small image pixel size and UAV body vibration interference, a novel aerial image object detection model based on the rotation-invariant deep denoising auto encoder is proposed in this paper. Firstly, the interest region of the aerial image is extracted by the selective search method, and the radial gradient of interest region is calculated. Then, the rotation invariant feature descriptor is obtained from the radial gradient feature, and the noise in the original data is filtered out by the deep denoising automatic encoder and the deep feature of the feature descriptors is extracted. Finally, the experimental results show that this method can achieve high accuracy for aerial image target detection and has good rotation invariance.

In view of the difficulty of noise removal in color images, according to the unique color separation charac-teristics of HSI space, the noise removal of color image polluted by high noise is studied. Firstly, the color image is projected into the HSI of the color feature space, and the color information is separated from the intensity feature information. Then, the color components H and S in the space are processed with a method of distance threshold denoising in polar coordinates, so as to remove the noise without losing the true color. At the same time, multi-scale transformation is carried out on intensity feature component I to obtain the high and low frequency subgraph. According to the frequent noise mutation in the high frequency subgraph, an adaptive gradient threshold denoising method is proposed to remove the noise in the high frequency to improve the image quality. The sparse denoising method is adopted to deal with a small amount of noise in the low-frequency subgraph. Finally, the corresponding inverse transformation is carried out to obtain the final color image. The experimental results show that the method in this paper achieves good denoising effect in vision and objective indexes such as PSNR, RMSE, SSIM and RE.

Pixel extraction accuracy is one of the key factors for geometric calibration accuracy of high resolution optical remote sensing satellites. A pixel extraction method is presented based on the light and small, automated reflected point source, which obtained pixel coordinates of point source image by parameterized Gaussian model fitting. The preliminary results show that the collinear error of pixel coordinates extracted by the reflected point source method is less than 0.001 pixels. The relative standard deviation of the proportional constant is also less than 5‰. And pixel coordinates extraction accuracy of point source image is less than 0.05 pixels comparing with common methods such as centroid method. The reflected point source can not only achieve high-precision pixel extraction, but also realize high-precision geometric calibration and image quality evaluation of high resolution optical remote sensing satellites, which is of great significance for improving the accuracy of China's space remote sensing stereo mapping.

Spectral library-based hyperspectral sparse unmixing technology has received attention in recent years, which uses spectral samples in the spectral library as endmembers and transforms the unmixing problem into a sparse representation problem. However, due to differences in the measurement environment, the actual endmembers of the hyperspectral image to be unmixed tend to differ from the corresponding spectral signatures in the spectral library. In this paper, an unmixing algorithm named spectral difference sparse constrained collaborative sparse regression is proposed. Firstly, we assume that the spectral differences have sparse property, and a spectral library correction model is established, which can make the spectral library be adaptively adjusted during the unmixing process; Then, the spectral library correction model is combined with the collaborative sparse regression unmixing model to establish a sparse unmixing model considering spectral differences; Finally, an iterative optimization solution based on the alternating direction method of multipliers is given. Synthetic and real hyperspectral data are used to verify the performance of different algorithms. The results show that the proposed algorithm is more effective than the compared algorithms in the presence of spectral library mismatches.

Hyperspectral remote sensing image has the characteristics of rich spectral information and combining image with spectrum, which has been widely applied in the earth observation. Most of traditional hyperspectral image classification models don't make fully use of spatial feature information, rely too much on the spectral imformation, making the classification accuracy still have a lot of room to improve. Conditional random field (CRF) is a kind of probability mode that can better integrate spatial context information. It plays a more and more important role in hyperspectral image classification. However, most CRF models have the problem of excess smoothness, which will result in the loss of detail information. Aiming at this problem, this paper proposed a hyperspectral image classification method based on space-spectral fusion conditional random field. The proposed method designs suitable potential functions in a pairwise conditional random field model, fusing the spectral and spatial features to consider the spatial feature information and retain the details in each class. The experiments on two sets of hyperspectral image showed that, compared with the traditional methods, the proposed classification method can effectively improve the classification accuracy, protect the edges and shapes of the features, and relieve excessive smoothing, while retaining detailed information.

To further improve the recognition accuracy of remote sensing image building area recognition, a classification method based on median robust extended local binary pattern(MRELBP) feature, Franklin moment and optimized support vector machine(SVM) by cuckoo search is proposed. Firstly, calculate the texture feature vector of the image block with MRELBP feature operator and use Franklin moment obtain the shape feature vector, the texture feature vector and the shape feature vector are combined into a comprehensive feature vector. Then train the SVM with training image samples, meanwhile, use cuckoo search to optimize the kernel function parameter as well as the penalty factor. Lastly, input the recognizing image into the SVM to get the result of building area recognition. The results of 30 groups of experiments show that, compared with the classification method based on RGB and SVM, the classification method based on LBP and SVM and the classification method based on Zernike moment and SVM, the accuracy of the remote sensing image building area identified by the proposed method is higher.

The development and characteristics of high-resolution optical/SAR satellites are introduced. Geometric radiometric calibration is an important means to improve the image quality. The basic principles of optical/SAR geometric radiometric calibration are summarized, and the research progress of geometric radiometric calibration is summarized. Finally, the frontier issues of satellite geometric radiometric calibration are summarized.

This paper presents a review of the advances, products and applications of lunar photogram-metric mapping using orbital images. It mainly introduces the key techniques such as construction and refinement of geometric models of lunar orbital images, co-registration of lunar orbital image and digital elevation model, optimal selection of multiple images, and large area mapping. Engineering and science applications of lunar photogrammetric mapping products are described, with more emphases on the Chang'E lunar exploration missions. Finally, the future research directions are discussed.

With the development of radar imaging technology and the diversification of imaging modes, the accuracy of coregistration in the direction of azimuth required by SAR interferometry is different. At present, due to the strict theoretical model and the high accuracy, the coregistration scheme with the image information based on geometric coregistration under the condition of precise orbits become the first choice for interference processing. In this paper, TerraSAR-X images have been taken as an example to demonstrate the accuracy of interferometric coregistration and satellite orbit required by different imaging modes. Furthermore, theoretical analysis and experimental validation have also been applied. It is proved that under the condition of precise orbits, SAR interferometric processing of strip-mode image of TerraSAR-X and other satellites can be achieved by geometric registration. While, spotlight imaging mode requires image coherence or spectral diversity to further optimization of the geometric coregistration results. Since the ESD (Enhanced Spectral Diversity) offset estimation method offers the highest accuracy. It has been used to estimate the geometric coregistration errors of Sentinel-1 satellite images with different orbit and DEM conditions. The results suggest that geometric coregistration error is mainly caused by the tangential error of the satellite orbit, and the difference of geometric corregistration of three kinds of DEM is much less than 0.001 pixels, which can meet the needs of Sentinel-1 image interference registration.

In order to improve the accuracy of identifying the target image while the drone is sailing, a hybrid model combining convolutional neural network (CNN) and probabilistic neural network (PNN) is proposed. The model uses CNN to extract multi-layer image representations and uses PNN extraction features to classify images. In order to improve the generalization and robustness of CNN, CNN-PNN model replaces BP neural network inside CNN with PNN, and trains the model by mean square error and gradient reduction method. The pre-processed image is transmitted to the CNN-PNN model to classify and identify the texture and contour of the image, and the simulation results of this model are compared with the results of convolutional neural network model and convolutional neural network-support vector machine model. The simulation results show that the CNN-PNN model has better accuracy compared with the two models, and the recognition rate is as high as 96.30%. The improved model improves the generalization and robustness of CNN, and can enhance the accuracy of image recognition in all aspects, and has high real-time performance.

Traditional aircraft detection algorithms which adopt handcraft features have poor performance in complex scene images and recognizing multi-scale objects. Methods using deep convolutional neural networks still face difficulty in dim small target search and recognition in large images with complex background. Aiming at these problems, a coarse-to-fine algorithm for aircraft detection in remote sensing images using cascade convolutional neural networks is proposed. To quickly and effectively acquire suspicious regions of interest (ROI), the whole image is searched by a small and shallow fully convolutional neural network which could deal with images of any size. Then deeper convolutional neural networks are used to refine the classification and location of the ROIs. A multilayer perceptron is introduced to the convolutional layer to improve identification capability of the convolutional neural networks and the strategies of multi-task learning and offline hard example mining are adopted in the process of training. At the detecting stage, the image pyramid is constructed and the redundant windows could be eliminated by the non-maximal suppression. Multiple datasets are tested and the results show that the proposed method has higher accuracy and stronger robustness and provides a fast and efficient solution for object detection in large remote sensing images.

Aiming at the problem that high-resolution optical remote sensing image lines are prone to fracture, a line segment optimization algorithm is proposed. Firstly, the line segment is regarded as a way to express the contour of the ground object, and the laws of line segment fracture from two aspects of geometric features and texture features are analyzed. Secondly, the line segment optimization algorithm is proposed. It takes the results of detecting line segments as the processing primitives, determines the initial optimized line segment according to the length of the line segment, establishes the tracking rectangular region and geometric constraint model for the broken line segments,build a dynamic optimization model, and give a complete line optimization process. Through the analysis of experimental results of some actual scenes and different types of remote sensing images, it is shown that the proposed algorithm can not only solve the problem of line segment fracture caused by terrain occlusion, edge blurring and edge serration, but also comparing with other methods, the proposed algorithm has great advantages in optimizing line length and restraining over extraction problem.

To solve the problem that when positioning the inner and outer edges of the iris, insufficient contour information may result in inaccurate positioning or failure，a positioning algorithm combining Hough transform and least squares method is proposed.Firstly morphological opening is used to reduce isolated small dots in the image,and adaptive threshold algorithm is used to convert images into binary images.Then the morphological closing operation is used to complement the small gap to smooth the edges of the image and Canny edge detection is used to extract image edge information.Finally, according to the difference between the inner edge and the outer edge, the inner and outer edges of the iris are positioned by Hough transform and least squares respectively.Verified by experimental comparison and analysis, this method can locate iris faster and more accurately than traditional algorithms.

Diffusion-weighted magnetic resonance images contain rich information on brain white matter (WM), and have been used to construct brain templates/atlases. However, the accuracy of brain templates reconstructed with the diffusion tensor model is often limited in regions with complex WM configurations. To overcome this problem, high angular resolution diffusion imaging acquisition and reconstruction have been developed for building higher quality brain template. In this paper, the research progresses on the construction of brain template using diffusion tensor imaging (DTI) are reviewed. Firstly, the technical aspects and limitations in using DTI data to build brain templates were discussed. Secondly, diffusion spectrum imaging and high angular resolution diffusion imaging techniques were introduced, with their advantages over the DTI techniques explained. Finally, the existing problems and the perspectives of the field were discussed.

In order to improve the problem of low accuracy in human behavior recognition task, a neural network based on batch normalization convolution neural network(CNN) and long short-term memory(LSTM) neural network is proposed. The CNN part introduces the idea of batch normalization, and the training data of the input network are normalized in mini-batch. After full connection, they are sent to long short-term memory neural network. The algorithm adopts the space-time dual stream network model structure. The RGB image of video data is taken as spatial stream network input, and the optical flow field image is taken as time flow network input. Then the recognition results obtained by the time-space dual-stream network are combined in a certain proportion to obtain the final behavior recognition result. The experimental results show that the space-time dual stream neural network algorithm designed in this paper has a high recognition accuracy in human behavior recognition tasks.

The goal of image coloring is to assign color to each pixel of the grayscale image, which is a hot topic in the field of image processing. U-Net is used as the main line network, and a fully automatic coloring network model is designed based on deep learning and convolutional neural networks. In this model, the branch line uses the convolutional neural network SE-Inception-ResNet-v2 as a high-level feature extractor to extract the global information of the image, and the Power Linear Unit(PoLU) function is used to replace the Rectified Linear Unit(ReLU) function in the network. The Experimental results show that this coloring network model can effectively color grayscale images.

A method is proposed for generating digital surface model (DSM) of high resolution satellite imagery (HRSI) based on modified semi-global matching (SGM) algorithm.Firstly,the system error of the rational function model is compensated by using the geometric constraint relation between the image connection points.Based on the compensation model,the image is divided into blocks.The projection trajectory method is used to obtain the image pairs of the images.In the dense matching stage,the disparity map is computed using semi-global matching by layer after building the pyramids images,and an expansion corrosion algorithm for disparity graphs,which takes into account the image texture information,is introduced to constrain the range of parallax search,increase the number of effective pixels at the edge of the parallax map and reduce the memory overhead and computation time required for the algorithm.In the post processing stage of disparity image,the edge information of disparity image is protected by weighted median filtering algorithm.Finally,the DSM is acquired based on the forward intersection.The stereo images of World View 3 and ZY-3 to experiment are selected.The experimental results show that the DSM accuracy obtained by this method is nearly 1.5 times higher than that of GSD in elevation direction,and the edge characteristics of the object are maintained well.The algorithm is computationally efficient and has relatively low memory overhead.

In the process of image retrieval, the traditional single feature can't reflect the distribution and details of image color and content, which have some adverse influence on the performance of image retrieval. This paper presents an image retrieval method based on improved color and texture feature. According to the mean of the HSV color model region, algorithm obtains the mean eigenvectors of the color feature by using the improved correlation weight model. The image decomposition transformation is obtained through the Haar wavelet. In the low-frequency component of the image decomposition, the low-frequency texture feature vector is obtained according to the low-frequency feature structure model. The similarity of image is calculated by the Canberra distance. Experimental results show that:the methods of retrieval are tested in Corel-1000 and Corel-5000 standard gallery, which accuracy rate and retrieval rate have been improved accordingly.

To solving the problems that the spectral threshold selection of image cloud detection and the influence of cloud-like ground objects on cloud detection results, a novel cloud detection method for HSRI based onthe spectrum and texture of objects is proposed.Firstly, histogram equalization is performed on the image, and then the appropriate image cloud detection spectral threshold is obtained according to the image equalization histogram.Secondly, the image is segmented to obtain superpixels using the simple linear iterative clustering algorithm.The cloud in the image is initially detected based on cloud detection threshold and spectral attributes of superpixels.Thirdly, the local binary patterns (LBP) texture image of histogram equalization image is obtained.The initial cloud detection image is refined based on the gray mean value and angular second moment of the superpixels LBP texture to eliminate the influence of cloud like objects.Finally, the cloud detection image is processed using region growing algorithm and expansion algorithm to obtain accurate cloud detection results.The experimental results show that the proposed method can obtain good cloud detection results.

Averaging of multiple scans is often used in magnetic resonance imaging (MRI) to increase the signal-to-noise ratio (SNR). However, image averaging often results in movement-induced blurs of the edges and tissue details. A matched and weighted averaging (MWA) method has been proposed by our group to obtain images with reduced blurring effects in signal averaging. Here a rotation-invariant non-local means (RINLM) algorithm was proposed, which used circular patches consisted of series of rings with equal area, instead of square patches, to search for similar patches in the images. Compared with the non-local means (NLM) algorithm, the RINLM algorithm was capable of finding more similar patches in the images containing many rotated local structure. This method was used to process noisy images to improve the SNR, and validated using both phantom images and in vivo MR images. The results demonstrated that the method could improve the SNR, while better preserving the edges and details of the images.

The traditional fuzzy clustering algorithm uses Euclidean distance as the similarity criterion, which is disadvantageous to the multidimensional data processing. In order to solve this situation, Mahalanobis distance is used instead of the traditional Euclidean distance, and the optimization of fuzzy clustering algorithm based on Mahalanobis distance is studied to enhance the clustering effect and ability. With making the initialization means by Heuristic search algorithm combined with k-means algorithm, and in terms of the validity function which could automatically adjust the optimal clustering number, an optimization algorithm KM-FCM is proposed. The new algorithm is compared with FCM algorithm, FCM-M algorithm and M-FCM algorithm in three standard data sets. The experimental results show that the KM-FCM algorithm is effective. It has higher clustering accuracy than FCM, FCM-M and M-FCM, recognizing high-dimensional data clustering well. It has global optimization effect, and the clustering number has no need for setting in advance. The new algorithm provides a reference for the optimization of fuzzy clustering algorithm based on Mahalanobis distance.

According to the existing Retinex algorithm can not automatically adjust the parameters, an adaptive bilateral filter Retinex algorithm based on parameter estimation was proposed. Firstly, the principal component analysis and the Canny edge detection algorithm are used to estimate the noise and edge intensity. Then the spatial geometric standard deviation parameter and the luminance standard deviation parameter of the bilateral filter are calculated by linear correlation calculation. And the bilateral filter of the parameter estimation is used to decompose the image and the reflected image. Finally, the illumination image and the reflected image are compressed and enhanced in different ways and a new image is synthesized. The experimental results show that the algorithm can not only automatically set the parameters, but also effectively inhibit the halo phenomenon.

The positioning accuracy of pipeline leakage mainly depends on denoising effect and time delay estimation performance. In view of the problem of poor positioning accuracy of current pipeline leak, Non-Local Means-Empirical Mode Decomposition(NLM-EMD) autocorrelation method for denoising and Fuzzy C-Means clustering(FCM)-second correlation method for adaptive time delay estimation are proposed. NLM is used for noise reduction pretreatment, and the signal is recombined and de-noised according to autocorrelation criterion of EMD, then FCM is used to adaptively extract the pressure drop signal with higher correlation, and second correlation can estimate the delay of the signal. The results show that NLM-EMD autocorrelation method can effectively reduce noise interference and improve the decomposition quality of EMD, FCM-second correlation method can enhance the correlation of the two leakage signals, and the positioning accuracy is improved by 0.817% compared with the direct second correlation.

A classification method of hyperspectral images based on deep 3D convolution networks is proposed in order to deal with the high dimensional and small samples of hyperspectral image classification. The method first uses hyperspectral data cube as input, and uses 3D convolution operation to extract 3D spatial-spectral features of hyperspectral data cube. Then, the residual learning is used to construct the deep network and extract higher level feature expression to improve the classification accuracy. Finally, the Dropout regularization method is used to prevent overfitting. Experiments were conducted on the University of Pavia, Indian Pines and Salinas datasets, and the results demonstrate that compared with support vector machine and the existing deep learning classification method for hyperspectral images, the method can effectively improve the classification accuracy of hyperspectral image.