Some ATR Technologies Based On Visual Attention Mechanism For IR Image Recognition

Automatic targets recognition in IR images is a hot research field of national defense technology. As to ATR in IR image, the interested target could be divided into two sorts: one is notable in shape, another is not. Examples for the first sort are airstrips, bridges and radomes, and for the second sort are the majority of mobile targets and depots of oil and ammunition. Recognizing targets of the second sort is a difficult problem in ATR field. The disadvantage of the conventional method to this problem is that it wastes too much time. Creatural visual attention mechanism is introduced into ATR field to conquer the disadvantage in recent years which performs well in experiment. The advantage of ATR based on visual attention mechanism is that it only processes the salient region in the image which may contain interested targets and the redundant information is left out of account ahead. Thus, this method consumes less time and performs better. This paper is composed mainly studies some ATR technologies for IR image recognition which based on these research achievements. Main contents of this paper are as follows:An approach to extract salient region based on local standard deviation is proposed after analyzing and studying the visual attention model earnestly. Local saliency of each pixel in the image is estimated by calculating its local standard deviation. And then, global saliency is estimated by appraising local saliency of each pixel in the global map. Lastly, visual salient map is obtained via syncretizing global saliency of different feature maps. The salient region can be detected and its dimension can be estimated in the map. This method is proposed based on the machine visual model proposed by Itti but is easier to be implemented than Itti's.An image segmentation method based on luminance and space information is proposed to solve the problem of the salient region segmentation. Segmenting the salient region by dint of its luminance information or space information is determined via judging the attribute of its histogram. The salient region then is segmented using the selected method. This method consumes less time and performs better than the conventional ones that is illuminated by the comparative experiment.The method to design classifier which divides features of the target into extra-category features and intra-category features is studied. The extra-category feature is the one which could distinguish one sort of targets from others while the intra-category feature is the one that all targets of the same sort take on. Firstly, we obtain extra-category features and intra-category features by practicing samples of the target. Then each feature is evaluated based on its reliability. As the extra-category features play major role in the process of ATR while the intra-category features play less important role, the former get higher scores and the latter get lower ones. All the scores will operate in the process of ATR as system parameters. The advantage of this classifier is that it makes the ATR system more reliable as different features play different role in the process of ATR based on their reliability. ATR based on score mechanism is proposed which selects feature parameters based on the application background. It firstly compares the feature of the target needed to be recognized with that of certain known target in the classifier, if two features match well, the kind-unknown target gets the score of this feature. The probability of the kind-unknown target belongs to the known kind is reflected by the scores. Finally, the classifier adjudges it belonging to the kind to which it gets highest score. The advantage of this ATR system is that it is adjustable between ATR precision and time consumption.