Research On Bionical Vision Model And Its Applications

“Knowing what is where by seeing” is the main task of human visual system. The study of physiology shows that there are two visual pathways, which process the visual task about “where” and “what”. The study of this dissertation is based on the mechanisms of two visual pathways, combines the research findings of brain science and cognitive science and also makes the biological mechanism as guide and restriction to take the visual modeling as a computational cognition problem. The research of this dissertation is to build the intelligent visual models and to solve the realistic problems.The contributions of this dissertation are as follows:Firstly, the dissertation proposes a novel biologically motivated model for outdoors scene classification. The model combines biologically inspired feature and cortex-like memory pattern. First of all, the biologically inspired gist feature is used to characterize the content of scene image. Then the Incremental Hierarchical Discriminant Regression tree is utilized to simulate the generation and recall process of human memory. The association between the gist feature and the scene label is established in an incremental way. A cognitive model of the world is constructed by real-time, online learning, and the new scene can be differentiated by way of reasoning. The experimental results indicate that the incremental model effectively improves the classification accuracy rates to nearly 100% and significantly reduces the training costs when compared with other biologically inspired feature-based approaches. The new scene classification system achieves state-of-the-art performance.Secondly, the dissertation proposes a novel framework that based on the proto-objects and visual memory mechanism to do saliency computations. There are two main tasks of saliency computation, the first one is “what factors affect saliency”.Based on the new physiology founding, the dissertation extends the concept of“proto-object” in computer vision. The semantic proto-object features, which involve all possible states of the observer’s memories such as face, person, car and text, are included in the concept of “proto-object”, as well as with the salient pixels, region, and the objects which are extracted in a bottom-up pattern. Experimental results indicate the effectiveness of the proto-object features.Thirdly, the dissertation proposes two methods to build “proto-object features-saliency” function. Two functions are built from two different ways to fuse the proto-object features and the top-down guidance. One of them is based on the “biased competition model”, and using the Gaussian processing to simulate the processing that memory affects the visual saliency; the other is based on the “feature-integration theory”, and the support vector machines are utilized to simulate the learning process.As a consequence, the association between the proto-object features and the salientinformation is established. Two visual attention models: POGP and POSVM are built via the method of machine learning, and the saliency information of a new image can be obtained by the way of reasoning. These two models have different applications. To validate the model, the eye fixations prediction problem on the MIT dataset is studied.Experimental results indicate that the proposed models effectively improve the predictive accuracy rates compared with other approaches.Fourthly, this dissertation proposes a manifold learning-based proto-object detection method, by studying the way that visual cortex process the visual information.It is believed that the changes of the local manifold lead to the change of smooth, and the changes correspond to the saliency of some samples. Therefore, in the image, the pixels that break the smooth of the manifold are salient. The value of salience depends on the level of destructiveness. At this point, a new method detects the salient region is based on the manifold learning which is proposed in a bottom-up way.Fifthly, this dissertation applies the bionical vision model to the work of intelligent inspection in the Smart Grid. On one hand, by studying on the needs of imaging of the transmission equipment, we adopt the visual saliency-based compressive sampling method to balance the reconstruction complexities and the quality of image. On the other hand, a novel method for insulators detection in the image of overhead transmission lines which based on lattice detection is presented in this dissertation. The method simulates the way that human visual system detects texture pattern and gets effective results. All the above works provide a new way for the intelligent inspection of the transmission equipment.In summary, this dissertation investigates two important issues in human visual systems. There are outdoor scene classification and visual saliency computation.Moreover, tentative studies have been carried out on the work of intelligent inspection in the Smart Grid. This dissertation demonstrates the feasibility and e ectiveness of the proposed bionical vision model. This will spark a great interest of research in the related communities in years to come.