A Clutter Rejection Technique for FLIR Imagery Using Region-Based Principal Component Analysis

Automatic target recognition (ATR) systems generally consists of three stages, as shown in Fig. 1 [1]: (1) a preprocessing stage (target detection stage) that operates on the entire image and extracts regions containing potential targets, (2) a clutter rejection stage that uses some sophisticated classi"cation technique to identify true targets by discarding the clutter images (false alarms) from the potential target images provided by the detection stage, and (3) a classi"cation stage that classi"es each target into one of a number of classes. ATR using forward-looking infrared (FLIR) imagery is an integral part of the ongoing research at the US Army Research Laboratory (ARL) for digitization of the battle"eld [2]. FLIR ATR, however, is a challenging problem, because of the highly unpredictable nature of thermal signatures. This high variability of target thermal signatures is due to several reasons, including meteorological conditions, times of the day, locations, ranges, etc. The high variability of target signatures, target obscuration, and clutter in the background make the target-detection stage produce a large number of false alarms. The false alarms produced by the target-detection stage must be discarded at the clutter rejection stage; otherwise, the recognition performed by the subsequent classi"cation stage would be unreliable, regardless of how good the classi"cation technique is. Therefore, it is highly desirable to develop a clutter rejection technique that can substantially reduce the number of false alarms produced by the detection stage, and that is the focus of this paper. We present a clutter rejection technique that uses region-based PCA. With this technique, we propose to categorize all target images by clustering together the target images with respect to their similar sizes and shapes in order to form a group. Each group is further divided in to several regions, and a PCA is performed for each region in a particular group to extract feature vectors. We propose to use feature vectors of arbitrary shapes and dimensions that are optimized for the topology of a target in a particular region. One can then use these feature vectors to decide whether a potential target is a clutter or a real target. The proposed technique is based on learning vector quantization that generates codebooks of the most representative feature vectors for each region in a particular group. The decision about a potential target is then made based on the similarity between the extracted feature vectors and the corresponding regional representative feature vectors, as well as the number of similar feature vectors found. In Section 2, we present the proposed technique and give experimental results in Section 3.