Accurate and Robust Automatic Target Recognition Method for SAR Imagery with SOM-Based Classification

Supervised SOM Based ATR Method with Circular Polarization Basis of Full Polarimetric Data

“…We have already proposed the efficient ATR method based on the supervised SOM and the U-matrix metric [3]. This section briefly describes the methodology of this method, because it is basis of our proposed method.…”

“…It has been reported that this method significantly enhanced the recognition accuracy compared with obtained by the neural network approach, by exploiting the feature of U-matrix field [3]. However, this method suffers from the degradation of recognition performance, where an unknown target image has a different azimuth angle of the training data.…”

“…However, such approaches result in seriously degraded classification accuracy when the available SAR images are highly contaminated by random noise. To overcome this difficulty, we have previously proposed the ATR method based on a supervised self-organizing map (SOM) method, where the unified distance matrix (U-matrix) metric is employed in the classification stage [3]. This method has been demonstrated to enhance the ATR performance remarkably compared with that obtained by major neural network or SVM-based approaches because it assesses not only differences in output from training and test inputs, but also the potential barrier generated by the U-matrix metric.…”

“…As to classification issue, this paper also extends the conventional SOM based ATR method [3] to be suitable for full polarimetric data, where the dimension of input vector is simply trebled. Furthermore, the conventional method suffers from inaccuracy for classification, where an unknown input image has different azimuth angle from the training image, namely, in target rotating situation.…”

An Advanced Neural Network-Based Approach for Military Ground Vehicle Recognition in Sar Aerial Imagery

“…The area of Automatic Target Recognition (ATR) for SAR imagery is an ongoing research in many branches of the military and large research institutions [6], [7], [16], [17], [20]. On the other side, there has been an increasing interest in using artificial neural networks (ANN) for image processing and pattern recognition [1], [2], [7], [11], [12], [13], [14], [15], [17]. A typical target recognition system consists of a detection module (filtering and segmentation) and a recognition module (feature selection and classification) [1].…”