Existing automatic target recognition of synthetic aperture radar (SAR ATR) schemes mainly focus on target chips, but there is very little research for a large-scale and high-resolution SAR image that is more practical for SAR image interpretation. How to recognize targets efficiently and accurately from a large-scale and high-resolution SAR image is still a challenge. We present a scheme based on the combination of a salient detection approach, an active contour model (ACM), an affine-invariant shape descriptor, and the corresponding shape context. During the detection stage, the spectral residual approach is utilized to efficiently preselect salient regions. The proposed convex ACM, based on a ratio distance and distribution metric which makes it more robust to multiplicative speckled noise, is then adopted to get accurate candidate target chips. For the discrimination stage, a cumulative sum of multiscale lacunarity feature is proposed to select vehicle chips from clutter chips. Finally, affine-invariant shape features, obtained from the contours by our proposed ACM, are combined with a corresponding shape context to make the classification more accurate. Experimental results demonstrate that our SAR ATR system, integrating all the proposed methods, is feasible in ATR from a high-resolution and large-scale SAR image.
The active contour model (ACM) is one of the most successful methods for target detection in optical and medical images, but multiplicative speckle noise greatly interferes with its use in synthetic aperture radar (SAR) images. To overcome this difficulty, a convex ACM is proposed. First, a ratio distance is defined in terms of the probability density functions on both sides of the contour. This ratio is then introduced into the Chan and Vese (CV) model. Second, by combining the modified CV and region-scalable fitting models, the global intensity fitting and local intensity fitting forces both evolve the contour. Third, the energy of the combination model is then incorporated into Chan’s global minimization active contour framework so that the global minimum can be reached. Furthermore, the gradient term in the final energy function allows the proposed model to locate boundaries more quickly and accurately. Numerically, a dual formulation is utilized to solve the problem of active contour propagation toward target boundaries. Target detection experiments in real and simulated SAR images show that the proposed model outperforms classical region-based and hybrid ACMs in terms of efficiency and accuracy.
Group tracking is to track groups of objects engaging in a common activity (formations), which is a challenging and significant problem for intelligent visual surveillance. It can develop an overall understanding in some applications. Also, information related to group behavior can be used to enhance the precision and reduce computational complexity for the estimate of individual target. In this paper a group initialization and tracking method is presented. Two stages are involved. First, groups are initialized by clustering detections of targets based on a mathematical morphological method. Second, the groups are tracked in terms of prediction, hypotheses, confirmation and deletion. Satisfactory results are shown when testing the algorithm on simulative sequences.
A novel algorithm for shape detection based on mathematical morphology is presented. Two stages are involved. In the first stage, a shape model is learned automatically from learning examples belonging to the same object class. It is a collection of subparts with the description of relations among subparts, represented by a fuzzy graph. In the second stage, the generated model is used to detect similar shapes from images of complex real scenes. Subparts of the shape are detected in sequence based on their saliency, and then the geometric configuration among those detected subparts is checked. A morphological component detector is proposed to detect each subpart by using a soft structuring element, derived from the shape model. Satisfactory results are shown when testing the algorithm on synthetic and real images.