P. Jarabo-Amores

Spatial-Range Mean-Shift Filtering and Segmentation Applied to SAR Images

The mean-shift (MS) algorithm is applied for reducing speckle noise and segmenting synthetic aperture radar (SAR) images. Two coastal images acquired by Envisats advanced SAR (ASAR) [European Space Agency (ESA)] are used. Studies of the MS parameters are carried out according to the desired product: a speckle filtered image where textures and edges are preserved, or a segmented image, where land and sea are distinguished, as a previous stage for obtaining a land mask and detecting the coastal line. In all cases, Gaussian kernels are used. Speckle filtering results are compared with those obtained using uniform kernels, proving that the former provides better results than the latter. A segmentation approach based on the positions and frequencies at which the MS converges is applied. The use of a combined spatial-range processing and the corresponding bandwidths makes the MS suitable for the two proposed problems. The solid theoretical basis of this procedure allows designing a guided search of the best parameters according to the desired solution, avoiding a tedious trial-and-error process. Although the used images have different characteristics, results prove that similar sets of parameters can be used, showing some degree of robustness with respect to the image, for a given sensor and image acquisition mode.

Approximating the Neyman-Pearson detector for swerling I targets with low complexity neural networks

This paper deals with the application of neural networks to approximate the Neyman-Pearson detector. The detection of Swerling I targets in white gaussian noise is considered. For this case, the optimum detector and the optimum decision boundaries are calculated. Results prove that the optimum detector is independent on TSNR, so, under good training conditions, neural network performance should be independent of it. We have demonstrated that the minimum number of hidden units required for enclosing the optimum decision boundaries is three. This result allows to evaluate the influence of the training algorithm. Results demonstrate that the LM algorithm is capable of finding excellent solutions for MLPs with only 4 hidden units, while the BP algorithm best results are obtained with 32 or more hidden units, and are worse than those obtained with the LM algorithm and 4 hidden units.

MLP and RBFN for detecting white gaussian signals in white gaussian interference

This paper deals with the application of Neural Networks to binary detection based on multiple observations. The problem of detecting a desired signal in Additive-White-Gaussian-Noise is considered, assuming that the desired signal samples are also gaussian, independent and identically distributed random variables. The test statistic is then the squared magnitude of the observation vector and the optimum boundary is a hyper-sphere in the input space. The dependence of the neural network detector on the Training-Signal-to-Noise-Ratio and the number of hidden units is studied. Results show that Radial Basis Function Networks are less dependent on the Training-Signal-to-Noise-Ratio and the number of hidden units than Multilayer Perceptrons, and approximate better the Neyman-Pearson detector.

Improving neural classifiers for ATR using a kernel method for generating synthetic training sets

An important problem with the use of neural networks in HRR radar target classification is the difficulty in obtaining training data. Training sets are small because of this, making generalization to new data difficult. In order to improve generalization capability, synthetic radar targets are obtained using a novel kernel method for estimating the probability density function of each class of radar targets. Multivariate Gaussians whose parameters are a function of position and data distribution are used as kernels. In order to assess the accuracy of the estimate, the maximum a posteriori criterion has been used in radar target classification, and compared with the k-nearest-neighbour classifier. The proposed method performs better than the k-nearest-neighbour classifier, demonstrating the accuracy of the estimate. After that, the estimated probability density functions are used to classify the synthetic data in order to use a supervised training algorithm for neural networks. The obtained results show that neural networks perform better if this strategy is used to increase the number of training data. Furthermore, computational complexity is dramatically reduced compared with that of the k-nearest neighbour classifier.

High order neural network based solution for approximating the Average Likelihood Ratio

The detection of gaussian signals with unknown correlation coefficient, ρs is considered. A strategy for designing high order neural networks (HONN) in composite hypothesis test is proposed. A HONN trained with ρs varying uniformly in [0, 1] is considered to approximate the Average Likelihood Ratio (ALR). In order to compare the suitability of the approximation, a sub-optimal solution based on constrained generalized likelihood ratio is used. A study of the computational cost is carried out. Results show that a HONN is able to approximate the ALR with a low computational cost.

Analysis of sea state parameters and ocean currents from temporal sequences of marine radar images of the sea surface

This work uses ordinary X-band marine radars to extract directional wave spectra and their related sea state parameters, as well as speed and direction of ocean surface currents, including tidal information. The used method analyzes the structure in frequency and wave number vector of the image spectra derived from of temporal sequences of marine radar images of the sea surface acquired by a marine radar system. The presented data and the related results were measured from a research platform located in the North Sea. In addition, the work presents some comparisons between sea state parameters derived from the marine radar analysis and the equivalent sea sate parameter obtained from in-situ wave sensor records.

Using multilayer perceptrons to align high range resolution radar signals

In this paper we propose the use of Multilayer Perceptrons (MLPs) to align High Range Resolution (HRR) radar signals circularly shifted in time. To study the performance, the error of shift estimation is measured for different values of Signal to Noise ratio (SNR). The Zero Phase method is used for comparison purposes. Results show the best performance of the Zero Phase method with completely misaligned patterns, and the best performance of the MLP with low grades of misalignment. Using these results, a new method is proposed. First, the Zero Phase algorithm is used to pre-align the signals. Then, a MLP is trained using the pre-aligned signals in order to get more accuracy on the estimation of the shift. Results show an improvement up to 30%.

Robustness of a Generalized Gamma CFAR ship detector applied to TerraSAR-X and Sentinel-1 images

A fast CFAR ship detector based on the statistical modeling of sea clutter in SAR images is proposed. Typical CFAR detectors, like the double parameter model (DPM), assume a Gaussian sea clutter model and usually degrade the image resolution using target windows. The proposed detector works in a pixel-by-pixel fashion, adaptively selecting a decision threshold for every single patch the SAR image is divided into. Therefore, the presence of different-sized ships will not become an issue and they will be better characterized, allowing the extraction of the target information, which could be used for refocusing, feature extraction, modeling and classification, while maintaining the resolution of the SAR image. Sea clutter is studied using different statistical models, with the Generalized Gamma distribution presenting itself as the most suitable one. This model is used to characterize the sea clutter in the proposed detector. In order to showcase the robustness of the proposed detector, images acquired with SAR sensors working in different frequency bands are selected. Ship detections results show a good performance regardless of the sensor, the ship size and the sea state.

Energy-weighted Mean Shift algorithm for speech source separation

Blind Source Separation algorithms have been applied to speech mixtures during many years, taking into account the knowledge and properties of speech signals. A new approach for speech separation based on sparse representations of speech has recently arisen. These methods are commonly known as Time-Frequency Masking methods, being the most famous the DUET algorithm that performs separation of undetermined mixtures from only two microphones. Sparsity property also encourages the idea of applying clustering techniques for source separation. In this work, we introduce an adapted version of the clustering method Mean Shift for the separation of speech sources. Obtained results confirm the validity of the method for speech separation improving the DUET performance and showing better generalization. Furthermore, the use of clustering techniques for separation enables the automatic identification of the number of sources.

Neural Solutions for High Range Resolution Radar Classification

In this paper the application of neural networks to Automatic Target Recognition (ATR) using a High Range Resolution radar is studied. Both Multi-layer Perceptrons (MLP) and Radial Basis Function Networks (RBFN) have been used. RBFNs can achieve very good results with a considerably small size of the training set, but they require a high number of radial basis functions to implement the classifier rule. MLPs need a high number of training patterns to achieve good results but when the training set size is higher enough, the performance of the MLP-based classifier approaches the results obtained with RBFNs, but with lower computational complexity. Taking into consideration the complexity of the HRR radar data, the choice between these two kind of neural networks is not easy. The computational capability and the available data set size should be considered in order to choose the best architecture. MLPs must be considered when a low computational complexity is required, and when a large training set is available; RBFNs must be used when the computational complexity is not constrained, or when only few data patterns are available.