Visual Cognition Based Recognition Of Objects In Natural Images

It is common knowledge that human beings have excellent ability of environmental perception and cognition, which enables us to easily identify the typical objects in the surroundings. With the increased understanding to the vision system of the human, how to improve the visual recognition ability of machines by applying information processing system of human visions has become a hot research topic in computer vision field. Yet the great challenge is how to imitate the typical function or information processing system of the brain vision system to equip the computers with the observing and understanding ability parallel to the mankind.This paper is oriented around the sub-topic of the National 973 Project– Key Technology of Environment Perception and Object Recognition Based on Vision Information (No. 2007CB31005). Based on the information processing system and principles of human retina and visual cortex, it establishes some cognitive computing models and methods with characteristics of human vision, realizes saliency processing of natural images and typical objects identification in natural environment, and provides calculable methods of visual environmental perception and information selective transmission system for automatic navigation of robots.Object identification in natural environment based on visual cognition is an emerging research subject, involving multiple disciplines, such as neurophysiology, cognitive psychology, biophysics, computer information, automation, etc. This paper makes some breakthroughs in both research method and design in this multidisciplinary research field, including six aspects as follows:①This paper makes systematic and all-around summary of functional experiment results for visual information processing made by neurologists at home and abroad. It analyzes and discusses different levels of functions in visual information processing. Special attention is paid to key technology and difficulties in biological-vision-based object identification in natural images.②This paper analyzes the information processing system of human retina in detail and presents a space-variant processing method for natural images. During the human visual perception process, the resolution of retina imaging varies with the changing spatial location of the focal point. Based on this fact, the paper simulates the perception system of human visual system. The research includes: (1) by simulating the perception feature of retina based on Gaussian multi-resolution pyramid and visual perception sensitivity to different dissolutions, a new image compression method is presented. In comparison with traditional uniform image processing approach, the presented method imitates non-uniform perception feature of biological retina; (2) through experiment, the basic data relationship between the location of retina fovea, image resolution and image compression rate is worked out, which provided reference to future research or engineering application. The presented method has valuable application potential in mass data remote transmission and image-based remote object identification.③Inspired from the environmental perception system of V1 of biological vision, this paper presents a contour extraction model and method for natural image centered on Gabor integrator. Researches by neurophysiologists show that V1 functions to weaken the environmental information and strengthen the repeating excitement at interest targets, so that it can extract the profile and boundary information of the object rapidly and efficiently, which provides important information for subsequent perception. This paper establishes target contour extraction method for natural scenes based on the analysis the information processing system of V1. This computing model can weaken irrelevant information while highlight target contour so as to extract the saliency of target profile. The research includes: (1) uniform Gabor integral kernel expression is established and its resemblance to biological cognition is proven mathematically; (2) the computing model and method with cyclical feedback, lateral inhibition, saliency strengthening is established, and natural image against complex background is processed to testify the feasibility and superiority of the presented method.④Inspired from visual"what pathway"information processing system, a natural scene target identification model and method based on perception invariance feature is established. Researches show that"what pathway"is the key information pathway for biological vision to identify objects. In this pathway, information processing begins from V1 area, goes through V2 and V4, and arrives at IT zone, where the high-level tasks of visual information processing and identification are finished. Based on this biological fact, a hierarchical computing model and method with"what pathway"functional features is established. This method can identify objects through effective extraction of object perception invariance and plastic learning. Experiments show that this method can effectively identify and classify typical objects in natural scenes, such as buildings, trees, sky, streets, pedestrians, vehicles, etc., with high robustness and recognition rate. ⑤Inspired from cognitive psychological research, this paper analyzes the weaknesses of traditional manifold cognition method (LLE) in characteristic learning and classification recognition, and improves the method with supervision. (1) Limit the searching scope of LLE data within the class, which can guarantee that only in-class data are used for restructuring weighting coefficients. In such way, computing resources can be saved while class feature can be highlighted. (2) After manifold feature learning, supervised second learning in Fisher subspace is conducted to low-dimension features, the learning results can maximize the inter-class dispersion and minimize the in-class dispersion of the sample data, so that the precision of classification is improved. In the experiment based on handwritten figures, this method shows the highest recognition rate.⑥Based on the non-Gaussian distribution feature of natural image data, this paper improves the autonomous mental development-based cognition method, i.e. ICS-HDR is presented on the basis of traditional HDR. This new method consists of three major steps: first, independent linear conversion of the sample image, converting high-dimension data to unrelated and independent components; then, feature selection based on reaction feature of biological visual cells to eigenvector dimensions; and at last, classification recognition to HDR. Compared with traditional methods, this improved method proves superiority in recognition rate and time consumption in both human face identification and obstacle identification in machine navigation.