Research On Key Technologies Of Scene Recognition Based On Biological Vision Mechanism

Scene recognition based on biological vision mechanism is defined as extracting scene characteristics in images by simulating the ability of human perception, deducing the class relationship among the images, recognizing the scene which the images belong to automatically. The scene recognition technology can do help with solving a set of representative application problems in the fields of computer science and pattern recognition, such as target tracking and positioning, video content analyzing, intelligent image retrieval, visual robot navigating and so on. For that, scene recognition has become one of the research issues which are very active and full of challenges, been widely concerned by researchers at home and abroad. Essential problems in scene recognition based on biological visual mechanism have been studied in this dissertation, the main research content and results are showing as following aspects:Because visual saliency detection can help to achieve the region segmentation with scene representation, study the attention selection mechanism in human visual system, compare the performance between models of spatial domain and frequency domain, visual saliency detection model based on frequency spectrum has been chosen, which has better performance. The feature and inner mechanism of detection model based on frequency spectrum, which is represented by SR, PFT and PQFT, have been analyzed. A unified framework of visual saliency detection algorithm based on frequency spectrum modulation has been proposed, and it contains all the algorithms above. On this basis, a new visual saliency detection model based on frequency spectrum balance modulation has been designed, which can solve the problem that traditional algorithm can’t detect precisely in particular condition. The simulation results show that the proposed ASBM model is superior to the PQFT model in detection accuracy, robustness and noise tolerance, and optimize the visual saliency detection algorithm.For the problem which is brought by the change of light to visual saliency detection and scene recognition, the image enhancement algorithm based on the theory of color constancy has been studied. The principle of multi-scale Retinex algorithm has been focused on, the shortcoming of the Gaussian Filter in the algorithm has been pointed out, then a multi-scale Retinex algorithm based on robust anisotropic diffusion has been proposed and the edge information which has the value of scene analysis has been protected to some extent and applied to color restoration. Comparison experiments show the proposed algorithm not only get better color restoration result, but also obtain clearer edge information, and provide essential security for region segmentation with scene representation and extraction of local invariant featureThe method of scene recognition based on image content characterization has been studied. A quick scene representative region segmentation framework based on visual saliency has been proposed, it can solve the problem of inefficient computation brought by traversing the whole image which is needed in traditional method of image processing in scene recognition to some extent. Among them, the region extracting algorithm based on entropy priority determines the center of scene representative region via computing the neighbor information of salient point, for that the algorithm has better robustness and noise tolerance, in the experiment the scene consistency between regions and local invariant interesting point has been proved good. While region segmenting algorithm based on priori knowledge can achieve precise detection of building regional information on image pixel level, the method has been applied to outdoor scene dataset. Compared with other methods, the segmentation algorithm not only detect the existence of saliency building, also extract the detailed regional information, and can remove the anti-occlusion and occlusion interference.The feature and building process of local invariant character descriptor have been studied, SURF algorithm with good performance has been chosen as a target to study and employ after comparison and analysis. A dominant direction navigation algorithm based on multi-direction integration has been proposed, it can obtain better dominant direction navigation result than the algorithm which exists in SURF. A competition strategy of similar interesting point based on the intensity of interesting points, contrast and density of interesting points of tiny-scale subspace has been proposed, which can remove noise points among the interesting points and improve the matching accuracy of local invariant character and recognizing efficiency of targets. The feature and flow of methods of scene recognition now have been studied, combining the stable scene recognition of small division size in this thesis, ISURF feature has been chosen for its better performance. For the problem of wasting computing source which is brought by global traversing, segmentation result of scene representative region gotten previously has been made use to reduce the to-be-searched space of ISURF feature. For the ownership of3layers’classes brought by image local characters, IHDR tree classifier based on scene analyzing mechanism and competing mechanism has been designed. Experiment show that the method of scene recognition based on ISURF feature and IHDR tree has been proposed with good recognizing results in accuracy rate and time-consuming.