Search

For hyperspectral image classification, this paper proposes a novel adaptive kernel sparse representation method based on multiple feature learning (AKSR-MFL). Firstly, multiple types of feature, including different kinds of spectral and spatial information, are extracted from the original HSI to describe the characteristics of pixels from different perspectives, which is beneficial to enhance the classification accuracy significantly. To further explore contextual information and conform the spatial structure as far as possible, we employ shape-adaptive algorithm to construct a shape-adaptive region for each test pixel at the same time. Then, we design adaptive kernel sparse representation (AKSR) method by applying kernel joint sparse pattern to address the linearly inseparable problem of classification in multiple feature space and make the pixels with the same distribution more easily grouped and linearly separable. Moreover, composite kernel constructed by multiple kernel learning (MKL) is embedded into AKSR to effectively construct base kernels for different feature descriptors and determine the weights of base kernels optimally, which can take the similarity and diversity of different types of feature descriptor into full consideration. Experimental results on three widely used real HSI data demonstrate that the proposed AKSR-MFL classifier outperforms several state-of-the-art classification methods.

Multispectral image compression can considerably reduce the volume of data and promote their application. However, conventional single-scale compression schemes, such as JPEG2000 and three-dimensional set partitioning in hierarchical tree (3D-SPIHT), do not accurately preserve the features of images due to the complex features of multispectral images. A compression framework based on adaptive multiscale feature extraction with a convolutional neural network is proposed. First, an adaptive multiscale feature extraction module, which is the basic component of the compression framework, is designed to extract the multiscale spatial–spectral features of the multispectral images and adaptively adjust the weights of the features according to the content of the images. Second, the encoder, which is composed of multiscale feature extraction modules, extracts the multiscale spatial–spectral features of the multispectral images, and the extracted features are quantized and encoded by the quantizer and the entropy coder to generate a compressed bitstream. Third, the decoder, which is structurally similar to the encoder, is utilized to recover the images. The rate-distortion optimizer is embedded in the encoder to control the trade-off between the rate loss and the distortion. The results of these experiments on multispectral images of the Landsat 8 satellite and the WorldView-3 satellite validate the better performance of our compression framework compared with the performances of conventional schemes, including JPEG2000 and 3D-SPIHT. In order to further verify the effectiveness of multiscale features, the framework is compared with a single-scale compression algorithm based on deep learning, the experimental results validate that the performance of the single-scale compression algorithm superior to the conventional schemes but inferior to our multiscale algorithm, which indicates that the multiscale features can significantly improve the performance of the compression algorithm.

We propose a compression framework based on multiscale features using deep learning technique for complex features of multispectral image. The multiscale feature extraction module is taken as the basic block of the framework to constitute the encoder and the decoder. The encoder is used to extract the dominant features of the input multispectral images and the decoder, which is symmetrical to the encoder, is used to reconstruct multispectral images. The balance of the bit rates and the distortion is implemented by the rate-distortion optimizer in loss function. The encoder and the decoder are trained jointly based on minimizing the mean squared error (MSE). Our proposed framework is compared with JPEG2000 in terms of PSNR and the experimental results validate that our proposed framework outperforms JPEG2000.

Aiming to the row sparse feature and the row sparse feature of the abundance matrix of hyperspectral mixed pixels, a hyperspectral unmixing method based on the non-convex sparse representation and spatial correlation constraints is presented. The non-convex sparse representation and spatial correlation representation models are firstly constructed, which take the non-convex p-norm of the abundance matrix to exploiting row sparse feature and the total variation regularization to exploit the spatial information of the hyperspectral images. Then the non-negativity constraint is taken into the model to exploit the physical property of the hyperspectral images. An iteratively optimization algorithm is developed for the unmixing model by the alternating direction method of multipliers. Experimental results illustrate that the proposed approach can get better unmixing accuracy than the other unmixing methods.

The autoencoder (AE) is based on reconstruction and unsupervised framework, so its extracted features contain enough components to represent the input signal. Based on this characteristic, AE can be well applied to hyperspectral unmixing. However, due to the low precision of traditional AE and the large influence of noise, in this paper, a deep denoising autoencoder network (DDAE) for hyperspectral unmixing is proposed to improve the accuracy of abundance estimation while realizing the de-noising function. In order to guarantee the abundance to be sum-to-one and nonnegative, the endmembers satisfy the nonnegative, we limit the weight of hidden layer and decoding layer to be nonnegative, add the sum-to-one constraint to the hidden layer, and the L2,1-norm constraint to the objective function as a regular term, which takes advantage of the multiple sparsity between adjacent pixels. Experiments with real data and comparison with other algorithms prove the effectiveness of the DDAE algorithm.

Considering that multispectral images have a mass of complex features, to extract intact feature data, a compression framework for multispectral images based on dense residual network (DRN) is proposed in this paper. The multispectral images are first fed into the encoder, and residual dense block in it can read all the information learned from the preceding block via a contiguous memory mechanism, then preserves the features adaptively. Then the data is compressed by down-sampling and converted to bit stream by quantization and entropy encoding. With that, global dense features can be extracted completely from the original image. Additionally, rate distortion optimization is used to make the data more compact. Then, we reconstruct the image via entropy decoding, inverse quantization, up-sampling, and deconvolution. The experimental result shows that the proposed method outperforms JPEG2000 at the same bit rate.

Recently, the rapid development of multispectral imaging technology has received great attention from many fields, which inevitably involves the image transmission and storage problem. To solve this issue, a novel end-to-end multispectral image compression method based on spectral–spatial feature partitioned extraction is proposed. The whole multispectral image compression framework is based on a convolutional neural network (CNN), whose innovation lies in the feature extraction module that is divided into two parallel parts, one is for spectral and the other is for spatial. Firstly, the spectral feature extraction module is used to extract spectral features independently, and the spatial feature extraction module is operated to obtain the separated spatial features. After feature extraction, the spectral and spatial features are fused element-by-element, followed by downsampling, which can reduce the size of the feature maps. Then, the data are converted to bit-stream through quantization and lossless entropy encoding. To make the data more compact, a rate-distortion optimizer is added to the network. The decoder is a relatively inverse process of the encoder. For comparison, the proposed method is tested along with JPEG2000, 3D-SPIHT and ResConv, another CNN-based algorithm on datasets from Landsat-8 and WorldView-3 satellites. The result shows the proposed algorithm outperforms other methods at the same bit rate.

Dictionary-aided unmixing has been introduced as a semi-supervised unmixing method, under the assumption that the observed mixed pixel of a hyperspectral image can be expressed in the form of different linear combinations of a few spectral signatures from an available spectral library. Sparse-regression-based unmixing methods have been recently proposed to solve this problem. Mostly, l p -norm minimization is a closer surrogate to the l 0 -norm minimization and can be solved more efficiently than l 1 -norm minimization. In this paper, we model the hyperspectral unmixing as a constrained l 2 , q – l 2 , p optimization problem. To effectively solve the induced optimization problems for any q ( 1 ≤ q ≤ 2 ) and p ( 0 p ≤ 1 ), an iteratively reweighted least squares algorithm is developed and the convergence of the proposed method is also demonstrated. Experimental evaluation carried out on synthetic and real hyperspectral data shows that the proposed method yields better spectral unmixing accuracy in both quantitative and qualitative evaluations than state-of-the-art unmixing algorithms.

3D motion estimation via feature point matching is an important issue in machine vision. However, most existing methods are not good enough due to various defects, such as being sensitive to matching error, and hard to give an appropriate threshold for feature point selection. To solve these problems, a mathematical model is obtained from massive experiments to describe how the quantity and quality of matched feature points influence the motion estimation accuracy. This model can be used to select the most appropriate feature points for motion parameter calculation, hence threshold setting is avoided while estimation performance is optimized. Experimental results show that the proposed algorithm can improve the accuracy of 3D motion estimation by up to 50% with little effect on computation time. Furthermore, it is especially suitable for real time applications as there is no need to set threshold manually.

The sparse representation method has been successfully applied in the field of hyperspectral image target detection. It assumes that target detection can be achieved by using target and background libraries to represent test pixels. Under this formulation, the presentation of the target and background signatures can be solved by L1-norm minimization of the weight coefficient and the target detection output is simply achieved by the difference between the two representation residuals. In this paper, a regularized multiple sparse Bayesian learning (RMSBL) method for hyperspectral target detection is proposed, which is established by Bayesian inference using the conditional posterior distributions of the model parameters under a hierarchical Bayesian model. According to the cost function for multiple sparse Bayesian learning, the presentation of the target and background signatures can be obtained by an L2,1-norm iterative minimization method. And the target detection result can be achieved with the difference between the two representation residuals. Four groups of hyperspectral datasets are used for simulation experiments. The results are compared with those of other common detection algorithms. The experimental results demonstrate that the RMSBL algorithm has higher detection performance.

Although spectral-spatial information can provide significant improvement for the accuracy of hyperspectral image (HSI) classification, it remains a challenging task to achieve better classification results with a small amount of training samples. To tackle this problem, we develop a novel nonlocal joint kernel sparse representation method based on local covariance (NLJKSR-LC) for HSI classification in this paper. Firstly, the maximum noise fraction (MNF) is applied to the original HSI to reduce the computational complexity. Also, several most similar local neighbouring pixels for each sample are refined from their corresponding superpixel to explore the local spatial information accurately. Then, we calculate the local covariance descriptor (LCD) among each pixel and its local neighbours, which can explore the spectral correlation of different spectral bands and offer important discriminative information for classification. Next, several nonlocal similar samples are searched from the whole HSI based on the nonlocal self-similarity property and a novel NLJKSR-LC model is established to address the problem of classification. The NLJKSR-LC model takes the local spatial context and the global structure of HSI into full consideration at the same time, which can improve the classification accuracy especially when there is a small amount of training samples. In addition, we design an appropriate kernel to address the problem that LCD lies in the linearly inseparable manifold space. Experimental results on three real-world HSI datasets demonstrate that the proposed NLJKSR-LC method outperforms the state-of-the-art classification methods with a small amount of training samples.

A regularized MSBL algorithm with total variation (RMSBL-TV) is proposed.Spatial and sparsity information are incorporated into the RMSBL-TV model.Propose an ADMM method to tackle the optimization problem of the RMSBL-TV model.We assess the performance of existing methods with the synthetic and real data. Sparse unmixing is a promising approach that is formulated as a linear regression problem by assuming that observed signatures can be expressed as a linear combination of a few endmembers in the spectral library. Under this formulation, a novel regularized multiple sparse Bayesian learning model, which is constructed via Bayesian inference with the conditional posterior distributions of model parameters under a hierarchical Bayesian model, is proposed to solve the sparse unmixing problem. Then, the total variation regularization and the non-negativity constraint are incorporated into the model, thus exploiting the spatial information and the physical property in hyperspectral images. The optimal problem of the model is decomposed into several simpler iterative optimization problems that are solved via the alternating direction method of multipliers, and the model parameters are updated adaptively from the algorithm. Experimental results on both synthetic and real hyperspectral data demonstrate that the proposed method outperforms the other algorithms.

Spatial information explored by superpixels can provide significant improvement for the accuracy of hyperspectral image classification. However, superpixels based sparse representation classification (SSRC) methods always introduce joint sparse pattern, which leads to the higher computational complexity. In this paper, a novel superpixel-feature-based multiple kernel sparse representation classification (SFMKSRC) method is proposed. Different from the traditional SSRC methods, SFMKSRC develops two new kinds of superpixel-based features by exploiting local mean operator within superpixels and weighted average operator among superpixels, which can describe the spatial information from both the local and global perspectives. Also, the novel superpixel-based features can avoid joint reconstruction to reduce the computational complexity. Then, the spectral-spatial features are extracted to combine with superpixels-based features and capture the detail structures of HSI to improve the classification performance. As pixels in high-dimensional feature space always trend to be linearly inseparable, a novel SFMKSRC model is designed to address the linearly inseparable problem of HSI classification. In addition, composite kernel constructed by generating base kernels for different features and optimally determining the weights of base kernels is embedded into MKSRC to exploit the strong correlations among different features while still preserve their diversities in a more flexible way. Experimental results on three real-world HSI datasets demonstrated that the proposed SFMKSRC method obtains a competitive performance and outperforms several state-of-the-art classification methods.

Deep learning has achieved great success in many computer vision tasks, such as recognition, image enhancement and image compression. However, it is difficult to use a residual network that can efficiently consider the characteristics of multispectral images for multispectral image compression. In this paper, a novel end-to-end multispectral image compression framework based on a weighted feature channel residual network is proposed to efficiently remove the spatial and spectral redundancy of multispectral images by extracting the importance of each channel. The multispectral image compression framework includes a forward coding network, a rate-distortion optimizer, a quantizer/inverse quantizer, an entropy encoder/decoder and an inverse decoding network. In the encoder, multispectral images are directly fed into the forward coding network, and the main spectral and spatial features of the multispectral images are extracted by the residual block. Additionally, the weighted feature channel module can explicitly model the relationship between feature channels when extracting features from multispectral images and adaptively allocate different weights for each feature channel through training. The rate-distortion optimizer is added to make the main features compact. Then, the intermediate feature data are quantized and encoded by lossless entropy coding to obtain a code stream. In the decoder, the code stream is approximately restored to the intermediate features through the entropy decoder and the inverse quantizer. Then, the intermediate features are reconstructed to the multispectral images by the inverse decoding network. Experimental results on 7-band multispectral images of the Landsat 8 satellite and 8-band multispectral images of WorldView-3 satellite demonstrate that the proposed algorithm can achieve a better PSNR than conventional 2D schemes (which are JPEG2000 and JPEG in this paper) and 3D scheme (which is 3D-SPIHT in this paper) and can effectively preserve more spectral information of multispectral images.

Based on the property analysis of interferential multispectral images, a novel compression algorithm of partial set partitioning in hierarchical trees (SPIHT) with classified weighted rate-distortion optimization is presented. After wavelet decomposition, partial SPIHT is applied to each zero tree independently by adaptively selecting one of three coding modes according to the probability of the significant coefficients in each bitplane. Meanwhile the interferential multispectral image is partitioned into two kinds of regions in terms of luminous intensity, and the rate-distortion slopes of zero trees are then lifted with classified weights according to their distortion contribution to the constructed spectrum. Finally a global rate-distortion optimization truncation is performed. Compared with the conventional methods, the proposed algorithm not only improves the performance in spatial domain but also reduces the distortion in spectral domain.

Sparse unmixing is a promising approach in a semisupervised fashion by assuming that the observed signatures of a hyperspectral image can be expressed in the form of linear combination of only a few spectral signatures (endmembers) in an available spectral library. Simultaneous orthogonal matching pursuit (SOMP) algorithm is a typical simultaneous greedy algorithm for sparse unmixing, which involves finding the optimal subset of signatures for the observed data from a spectral library. But the numbers of endmembers selected by SOMP are still larger than the actual number, and the nonexisting endmembers will have a negative effect on the estimation of the abundances corresponding to the actual endmembers. This paper presents a variant of SOMP, termed backtracking-based SOMP (BSOMP), for sparse unmixing of hyperspectral data. As an extension of SOMP, BSOMP incorporates a backtracking technique to detect the previous chosen endmembers’ reliability and then deletes the unreliable endmembers. Through this modification, BSOMP can select the true endmembers more accurately than SOMP. Experimental results on both simulated and real data demonstrate the effectiveness of the proposed algorithm.