Research On Object Detection And Recognition Inspired By Biologically Visual Perceptual Theory

Human visual perception system can easily perceive still or moving targets from the scene, which is very diffcult for computer devices. As the research focus in the computer vision domain, object detection and recognition attacts lots of research interest, which possesses lots of utilization potentiality. Until now, the visual perception ability of machines lags far behind human’s. With this consideration, designing the object detection and recognition algorithm through imitating the perception process wihch is reported in the latest research achievement in neurosciene and cognition, can be a very promising way to largely improve the object detection and recognition performance.Inspried by the human visual perception mechanism, this paper implements a series of object detection and recognition researches, and gives a focus on spatial salient object detection, spatio-temporal salienct object detection, infrared moving targets detection from airborne imagery, and unsupervised multi-level feature extraction and its application in the object classification technique. The aforementioned researches are specifically depicted in the following:Firstly, inspired by the attention enhancement theory, this paper proposes a Cauchy graph embedding based spatial salient object detection model. The enhancement theory shows that the object based attention result is the diffusion result of the space based attention result. Based on this theory, the given approach takses the eye fixation result as the input, proposes a novel Cauchy graph embedding based smoothing approach for implementing visual organization to imitat the attention diffusion process. Large amouts of experiments demonstrate that the Cauchy graph embedding based visual organization can outperforms traditional Laplacian graph embedding, and can obviously improves the salient object prediction performance. Notably, the final result can achieve the comparable result with the state-of-the-art approaches.Secondly, inspired by the two separated perception streams, this paper proposes a spatio-temporal salient object detection algorithm using the mixed feature based kernel regression. Two streams hypothesis is widely adopted: the visual perception process is first along the ventral stream and the dorsal stream, and is finally fused as a whole. Inspired by this visual perception process, the proposed spatio-temporal salient object detection approach is mainly composed by three modules:the spatial saliency detection module corresponding to the appearance perception function, the temporal saliency detection module corresponding to the motion perception function, and the spatio-temporal saliency fusion module corresponding to the information integration function. In order to cope with the fusion problem, this paper proposes kernel regression in mixed feature space (KR-MFS) for the first time. In addition, KR-MFS includes three entity estimators:kernel regression using local constant approximation in mixed feature space (KR-LC-MFS), kernel regression using local regularized linear approximaiton in mixed feature space (KR-LRL-MFS), and kernel regression using local regularized kernel approximation in mixed feature space (KR-LRK-MFS). Using KR-MFS, we propose a constructive hybrid fusion approach. Experiments show that the proposed hybrid fusion approach can outperform the existing fusion approaches remarkably. Benefiting from the hybrid fusion approach, the proposed spatio-temporal saliency model can outperform the state-of-the-art approaches.Thirdly, inspired by the feedback modulation, this paper proposes a multi-level infrared moving targets detection approach from the airborne imagery. Large amounts of neuroscience experiments show that the process and acitivation speed of the dorsal stream is much higher than the ventral stream. In addition, the fact that the ventral stream is modulated by the dorsal stream in a feedback fashion indeed exists. Inspired by this asymmetric modulation mechanism, we propose a real-time infrared moving targets detection approach from the airborne imagery. The proposed approach mainly includes three level steps:fast motion perception, coarsely motion focuse and views extraction, and accurately object detection in motion views. Large experiments show that the the proposed multi-level approach can outperform the existing methods in terms of efficiency and accuracy.Finally, inspired by the hierarchical perception mechanism of the visual cortex, this paper proposes an unsupervised multi-layer feature extraction method. Lots of neuroscience evidences demonstrate that human possesses the superior ability of object detection and recognition because visual cortex can automatically perceive the features from low to high. More specifically, the primary visual cortex (VI) is sensitive to the Gabor filter like edge response, and the higher visual cortex (e.g., V2) starts to be sensitive to more complex stuctures (e.g., corners, junctions and so forth). From the visual result, the basises from the first layer of the proposed approach are similar to the response of the cells in V1, and the basises from the second layer of the proposed approach are similar to the response of the cells in V2. In addition, the proposed feature extraction approach is tested in the object classification test dataset (i.e., CIFAR-10) and the scene classification test dataset (i.e., UCM). In the aforementioned classfication tasks, the proposed approach can achieve ideal classfication performance.