The algorithm of synthetic aperture radar automatic target recognition (SAR-ATR) is generally composed of the extraction of a set of features that transform the raw input into a representation, followed by a trainable classifier. The feature extractor is often hand designed with domain knowledge and can significantly impact the classification accuracy. By automatically learning hierarchies of features from massive training data, deep convolutional networks (ConvNets) recently have obtained state-of-the-art results in many computer vision and speech recognition tasks. However, when ConvNets was directly applied to SAR-ATR, it yielded severe overfitting due to limited training images. To reduce the number of free parameters, we present a new all-convolutional networks (A-ConvNets), which only consists of sparsely connected layers, without fully connected layers being used. Experimental results on the Moving and Stationary Target Acquisition and Recognition (MSTAR) benchmark data set illustrate that A-ConvNets can achieve an average accuracy of 99% on classification of ten-class targets and is significantly superior to the traditional ConvNets on the classification of target configuration and version variants.

Target Classification Using the Deep Convolutional Networks for SAR Images

Texture-based segmentation of synthetic aperture radar (SAR) image is a difficult task in remote sensing applications because it must address the problem of speckle noise. Several methods have been proposed for this purpose based on clustering, but suffer from long run times, computational complexity, and high-memory consumption. The proposed technique consists of two phases for SAR image segmentation. A new algorithm for parameter estimation based on curvelet coefficient energy (KCE) to design an optimum kernel function and an unsupervised spectral regression (USR) method have been proposed in phases 1 and 2, respectively, for SAR image segmentation. Eigen-decomposition is not required in USR, which decreases run times over other methods. The proposed algorithm uses a single-stage curvelet to extract the texture feature. Then, a new term is introduced based on the kurtosis feature value of the curvelet coefficients energy of the SAR image. Finally, the level set method is used to outline the boundaries between textures. Subimages are then extracted from the textures. After the Gabor filter is applied and the features are extracted, they are learnt using USR and clustered using a $k$ -means algorithm. It is demonstrated that the clustering results based on the learned features will be improved significantly. SAR image segmentation is performed using the $k$ -means after applying the Gabor filter bank and feature extraction. The results of segmentation are compared with Nystrom and parallel sparse spectral clustering (PSSC). The proposed method was shown to be more accurate and had a shorter run time than either Nystrom or PSSC.

A Two-Phase Algorithm Based on Kurtosis Curvelet Energy and Unsupervised Spectral Regression for Segmentation of SAR Images

To achieve efficient ship detection in high-resolution synthetic aperture radar images, an improved superpixel-level constant false alarm rate (CFAR) detection method is proposed with three modifications. First, the weighted information entropy (WIE) describes the statistical characteristics of superpixels, yielding a better distinction between target and clutter superpixels. Second, a two-stage CFAR detection scheme is proposed to detect target superpixels, including global detection and local detection. Specifically, the WIE-based global detector is utilized to prescreen candidate target superpixels (CTSs) and then the local CFAR detector is adaptively conducted over the selected CTSs to refine target superpixels. Third, the superpixels selected by the global detection can be excluded as outliers for clutter estimation, thus the disturbance from adjacent targets in multitarget situations can be reduced and more accurate detection results can be obtained. Compared with CFAR methods implemented with the sliding window technique, the computational burden of the proposed method is significantly reduced without loss of detection performance. Experiments of ship detection on three TerraSAR-X datasets are presented to validate the effectiveness of the proposed method.

An Improved Superpixel-Level CFAR Detection Method for Ship Targets in High-Resolution SAR Images

A new and robust constant false alarm rate (CFAR) detector based on truncated statistics (TSs) is proposed for ship detection in single-look intensity and multilook intensity synthetic aperture radar data. The approach is aimed at high-target-density situations such as busy shipping lines and crowded harbors, where the background statistics are estimated from potentially contaminated sea clutter samples. The CFAR detector uses truncation to exclude possible statistically interfering outliers and TSs to model the remaining background samples. The derived truncated statistic CFAR (TS-CFAR) algorithm does not require prior knowledge of the interfering targets. The TS-CFAR detector provides accurate background clutter modeling, a stable false alarm regulation property, and improved detection performance in high-target-density situations.

Robust CFAR Detector Based on Truncated Statistics in Multiple-Target Situations

To eliminate the influence of target returns on the estimation of local sea clutter distributions, an improved iterative censoring scheme (ICS) for constant false-alarm rate detectors is proposed with two modifications. First, the proposed ICS censors out both target pixels and their four-connected neighborhood pixels from the estimation of local sea clutter distributions. Second, a novel initial detector is proposed to improve the convergence speed of ICS. The proposed initial detector, which only needs the probability of false alarms as an input parameter, is based on the sorting of all pixels under test. Experiments of ship detection with RADARSAT-2 ScanSAR wide mode images are presented to illustrate the effectiveness and improvements of the proposed ICS.

An Improved Iterative Censoring Scheme for CFAR Ship Detection With SAR Imagery

In this letter, a censoring scheme that iteratively updates the outlier/target maps for target detection in synthetic aperture radar (SAR) images is proposed. For each iteration, any pixels that are indicated by the outlier map as outliers are rejected (censored out) from the clutter estimation. The resulting detected target map is then used as the new outlier map for the next iteration. This procedure is continued until there is no change to the target map, which is then output as the final detection result. The proposed scheme is generically applicable for target detection in both single-channel and multichannel SAR images. In our experiment, in particular, we tested the proposed method on both single-channel and polarimetric SAR data, and its effectiveness was successfully demonstrated.

On the Iterative Censoring for Target Detection in SAR Images

Recently, the use of linear features for processing remote-sensing images has shown its importance in applications. Unfortunately, traditional linear feature detection methods rely heavily on the image's local information which makes them vulnerable to the presence of noise in the image. This problem becomes particularly difficult for synthetic aperture radar (SAR) image applications where SAR images are corrupted by speckle noise. In order to overcome this problem, we propose a novel method that processes the polarimetric synthetic aperture radar (Pol-SAR) images by combining the multiscale image analysis with polarimetric information in a new fashion. A two-scale approach is adopted here. On a coarse level, the coarse regions of the linear features are extracted by a curvelet transform from a speckle noise reduced image obtained by the polarimetric whitening filter. On a fine level, we develop a fuzzy polarimetric detector to accurately locate the linear features inside the regions. The effectiveness of the proposed method is demonstrated using simulated Pol-SAR data acquired from both EMISAR and Convair-580 systems.

Linear Feature Detection in Polarimetric SAR Images

In this letter, an unsupervised method for estimating the equivalent number of looks (ENL) in synthetic aperture radar (SAR) images is proposed. Assuming that the multiplicative noise in the SAR images is a unit mean random variable of Gamma distribution parameterized by the ENL, a lookup table between the ENL and the variance of the logarithm of the multiplicative noise is established. Then, for a SAR image, we first apply the logarithmic transform to make the multiplicative noise become additive. After that, the variance of the additive noise in the logarithmic image is estimated; finally, the ENL is obtained according to the estimated noise variance by the pre-established lookup table. The effectiveness of the proposed method is demonstrated with both simulated and real SAR images.

Unsupervised Estimation of the Equivalent Number of Looks in SAR Images

Standardized use inspection of workers' personal protective equipment based on deep learning

In this paper, a new method for supervised classification of terrain types in polarimetric Synthetic Aperture Radar (Pol-SAR) images is proposed. This technique is a combination of the texture classification and the maximum likelihood classification based on the complex Wishart distribution for the polarimetric covariance matrix. The texture features are first extracted from the span image based on co-occurrence matrices; and then the classifier combines the texture features with the distance measure based on polarimetric information to obtain the results. Using a NASA/JPL AIRSAR image, the effectiveness of the proposed method is demonstrated.

Guangyi Zhou

Papers

On the Iterative Censoring for Target Detection in SAR Images

Linear Feature Detection in Polarimetric SAR Images

Unsupervised Estimation of the Equivalent Number of Looks in SAR Images

Standardized use inspection of workers' personal protective equipment based on deep learning

Pol-SAR images classification using texture features and the complex Wishart distribution