Statistical Analysis And Modeling Of Human Visual Cognition Patterns

As an important branch of Neuromorphic computing, research on the mechanismsand computational models of visual cognition and decision-making can not only bringscientific and technological advancements to research fields like neuroscience, computervision, artificial intelligence etc., but also provide strong fundamental support for nationalconstruction in the area of bio-medical system, intelligence services, military equipmentR& D etc. This thesis associates the theory of judgment and decision-making underuncertainty from cognitive psychology with specific visual computing problems in visualcognition research, and systematically investigates the computational models of severalkey factors influencing subjective judgments and decision-making during visual cognitionprocess. The proposed models were applied and evaluated comprehensively in real-worldapplications. The main contents and the contributions of this thesis can be summarized asfollows:Firstly, human eye fixations are analyzed from a pure statistical point of view, basedon which a novel saliency prior, namely the Super-Gaussian prior, is achieved. A dy-namical cognitive framework is then constructed which functions similarly with humansaccadic behavior. Experimental results including the responses to synthesised images,prediction of eye-fixations, as well as proto-object detection, demonstrate that the pro-posed model can outperform all the state-of-the-art attention(saliency) models on wellacknowledged benchmarks. Our model is able to rapidly discover salient patterns in vi-sual signals and effectively predict and simulate the eye movements of human subjects.The research of this part answered the question of “Which part of the visual signals is ac-tually available in the decision-making process?”, which corresponds to the “availability”heuristic in Kahneman’s theory.Secondly, this thesis studies several key issues of visual attention models and salien-cy detection algorithms. In model adaptivity experiment, similar image-level perfor-mances among the tested models have been observed, and the degeneration caused bylow inter-viewer inconsistency was also discovered. The experiments on the effects ofscale show that multi-scale strategy performs better in scenes containing salient objectsover multiple scales. Experiments on the effects of feature dimensions show that increas-ing the number of features might not always improve the model’s overall performance andusing a complete feature representation is usually not the optimal solution. Based on 20 computational models, a generic model boosting algorithm and a statistical multi-modalfusion strategy are also proposed.Thirdly, inspired by the Bayesian Set representativeness model, we integrated pro-totype theory in cognitive psychology with ontological knowledge bases from the field ofWeb knowledge mining research, and constructed customized image ontology to approx-imate the semantic context for the given target concept. Meanwhile, a dynamic clusteringalgorithm is adopted to discover the potential proto-types and obtain a compact repre-sentation for each related concept. Based on such implicit semantic knowledge, a rep-resentativeness model named Ontological Prototype Model is proposed. Compared withstate-of-the-arts, the proposed model can better characterize the representativeness of thesamples given a specific semantic concept. Results of image ranking on Image Net and ourWeb image database demonstrate the superior performance of the proposed model overseveral state-of-the-art approaches. The research of this part answered to the question of“Which part of the visual signal is representative for a given concept”, which correspondsto the “representativeness” heuristic in Kahneman’s theory.Finally, we explore the distribution bias of visual media on Web, and propose a nov-el hypothesis which associates the distribution bias of Internet images with the cognitivepreference of human visual perception. Based on such hypothesis, we construct our cog-nitive concept model which not only reflects the characteristics of Web image distributionbut also agrees with the cognitive behaviors of human beings. Experimental results showthat the proposed model is able to organize the visual media according to the related se-mantic clues, and can effectively extract the embedded cognitive patterns of human users(the anchors) from the Web, which could further establish the quantitative associationbetween general semantics and the visual media. The results indicate that there existslarge amount of discoverable human knowledge on Web, and they also reveal the formalmathematical connection between two independent psychological research topics: “thecanonical view of object” and “the representativeness heuristic”. The research of this partanswered to the question of ”Which part of the Visual signals is the anchor for humandecision-making”, which corresponds to the Kahneman’s “anchoring” heuristic.Through the above studies, this thesis deeply explored judgment-oriented cogni-tive mechanisms and the corresponding computational models. Experimental resultsshow that Kahneman’s theory on judgment under uncertainty can be naturally extendedto visual decision-making process. The proposed computational models for “availabili-ty”(Saliency), “Representativeness” and “Anchoring” heuristic can not only provide plau-sible explanations for classic human cognitive behaviors like saccadic eye-movements andcanonical views of objects, but also solve practical problems in real-world applicationssuch as object detection, image ranking, and iconic sample discovery.