Visual Saliency Detection With Its Applications In Automatic Target Recognition Systems

Visual attention mechanism reveals the visual perception mode of human eyes for input scenes. Visual saliency, with its capability to locate important regions covering valuable contents in an image, contributes to rapidly and accurately extracting effective information while discarding redundant information from complex scenes, which promotes to optimize the information processing efficiency in human brain. Its strong information filtering ability could not only help to save computing resources, but also improve the accuracy and robustness for crucial techniques including target detection, recognition and tracking in automatic target recognition (ATR) system, which is very instructive in developing ATR techniques towardsing intellectualization.Based on the discussion on the significance for studying visual saliency with its applications in ATR system, in this dissertation, including eye fixation prediction, salient object detection, saliency based small detection in infrared images and ship detection in remote sensing optical images are deeply explored. Specifically, several new algorithms are proposed, and the main work is listed as follows:For the research on eye fixation, it could find the foundation from recent investigations in primary visual cortex by discovering that the responses of neurons carrying effective information are sensitive and weakly correlated while neurons carrying redundant information are strongly correlated. In this sense, we propose a new visual saliency detection model by using feature activity weighted decorrelation cues, to validate the effectiveness of feature decorrelation and feature activity for predicting eye fixation. Experimental results demonstrate that the proposed model could obviously improve the prediction accuracy, which further provides some available achievements in this field.For the research on salient object detection, regarding the fact that traditional contrast based method is often invalid when there are complex clutters in background regions or big feature variances within foreground regions, we propose a new salient object detection method by combining contrast information and object vision organization cues. It consists of two stages:contrast saliency computation and object vision saliency filtering. The former addresses the uniqueness and compactness with background prior and center prior, and the later describes the properties of closure, proximity and similarity inspired by the Gestalt principles of object grouping. The proposed method is not only effective for complex scenes, but also applicable when the number of salient objects is unlimited.Motivated by Boolean map visual theory, a new single-frame based infrared small target detection method is proposed. Boolean map visual theory reveals that an observer’s visual awareness corresponds to one Boolean map via a selected feature at any given instant. According to the fact that small targets in infrared images are bright and have Gaussian-like shape in the local context area, we separate the image into a color channel and an orientation channel, and then combine all the weighted Boolean maps produced from these two feature channels, to hightlight targets and suppress backgrounds. This method could be widely applied in various scenes, furthermore, it is very robust for images with complex backgrounds.Inspired by the fact that visual saliency could highlight unique patterns from generic patterns, a new directional saliency-based method is proposed for infrared small target detection. It takes advantage of the texture differences between small targets and backgrounds, namely the target has isotropic Gaussian-like shape whereas background clutters are generally local orientational, so their textures could be seemed as two kinds of different patterns. Motivated by this observation, we use visual saliency to detect small targets in infrared images. Experimental results demonstrate that the proposed method could effectively overcome the backgrounds clutters for various complex scenes, and it has good performance for small target detection.We propose an unsupervised ship detection method from optical satellite images based on visual saliency and S-HOG descriptor. It is robust for scenes with cloud, wave and wake clutters, and is effective when ship size varies. In the proposed method, visual saliency contributes to extracting candidate regions and suppressing backgrounds, by highlighting salient signals while filtering redundant information. S-HOG is able to characterize ship targets with various sizes, which could overcome wake clutters as well. Experimental results demonstrate that our method is very effective for ship detection in remote sensing images.