Robust Infrared Superpixel Image Separation Model for Small Target Detection

Abstract

Accurate and rapid detection of small targets against complex background is a fundamental requirement of various computer vision systems. This work is the first attempt to apply the superpixel segmentation technology to the field of low resolution infrared small target detection in the extremely complex backgrounds. The main contributions are as follows. First of all, the simple linear iterative cluster (SLIC) algorithm is utilized to accurately classify the raw infrared image into three components: outlier superpixels, stable background superpixels and target superpixels, which appropriately aggregate similar background components as the basic unit of subsequent processing. In SLIC, an optional range of superpixels numbers is specified to robustly implement superpixel segmentation strategy on low resolution infrared images. Second, an outlier superpixel masking (OSM) model is proposed to perform accurate identification and masking of the outlier superpixels with highly heterogeneous backgrounds, thus minimizing false alarm rate. Specially, a 2D Gaussian filter matching the target distribution is introduced to blur the remaining boundary and pixel-sized noises with high brightness (PNHB) while maximizing the signal-to-noise ratio (SNR). Finally, a singular value truncation strategy with entropy weighted sparse factor (SVT-EW) is proposed to implement the final target extraction, which assigns specific sparsity weights for small infrared targets. SVT-EW effectively resolves the background residuals in gray-based threshold segmentation, and therefore generates precise target detection results. Extensive experimental results on fourteen extremely complex infrared natural scenes validate the superiority of the proposed method over the state-of-the-arts with respect to robustness and real-time performance.

PixelGame: Infrared Small Target Segmentation as a Nash Equilibrium

“…The traditional methods rely on hand-crafted features. They mainly either simplify the small target as a bright spot [7], [8], or model the background, target and the relationship between them [9], [10] in a particular scene. Most traditional methods are designed with specific yet limited features, which failed to cover multiple scenarios and thus resulted in degraded performance in open and diverse environments.…”