Cognitive and Neural Modelling for Visual Information Representation and Memorization

Limiao Deng

• Publisher: CRC Press
• Year: 2022
• Print ISBN:
  9781032249117
• eBook ISBN:
  9781000574654
• Topic: Computer Science
• Subtopic: Computer Vision & Pattern Recognition
• Index: Computer Science

Focusing on how visual information is represented, stored and extracted in the human brain, this book uses cognitive neural modeling in order to show how visual information is represented and memorized in the brain.

Breaking through traditional visual information processing methods, the author combines our understanding of perception and memory from the human brain with computer vision technology, and provides a new approach for image recognition and classification. While biological visual cognition models and human brain memory models are established, applications such as pest recognition and carrot detection are also involved in this book.

Given the range of topics covered, this book is a valuable resource for students, researchers and practitioners interested in the rapidly evolving field of neurocomputing, computer vision and machine learning.

CHAPTER 1 Introduction

DOI: 10.1201/9781003281641-1

1.1 Background

Humans can easily detect and recognize objects, no matter how different their appearance is and how complex the environment is; however, this basic function is a great challenge for computer vision systems. Humans can do this easily because they have a powerful system of visual perception and memory. Among the external information the brain receives, visual information accounts for a large proportion, about 80% of all perception information. The processing process in the visual system is very complex. Memory is very important for human beings. Because of memory, human beings have the ability to learn and remember what they have seen and experienced, which is conducive to rapid recognition of remembered scenes, cognition of new things, and adaptation to new environments. The main purpose of computer vision research is to enable computers to perceive, interpret, and understand the environment just as people do. However, the current level of computer vision is far from human vision. So, it is of important theoretical and applicational value to simulate the visual information processing of visual perception and brain memory mechanisms to achieve effective representation of visual information and rapid memory and to achieve high reliability, strong adaptability of image recognition and classification systems on the basis of the cognitive neuroscience research of visual perception and the brain memory mechanism¹.

How to process large amounts of visual information and extract useful information, and how to organize and represent the information effectively, have been a hot and difficult topic in the field of computer vision. The biological visual system shows excellent performance in this aspect. For example, in object recognition, no matter how the environment and light change, whether the object changes in scale, angle of view, rotation, position, and so on, the visual system can easily acquire the invariant characteristics of describing the object and perform rapid and effective recognition. Based on this, many researchers have tried to simulate the structure and function of visual cortex of the brain and constructed visual perception models based on biologically inspired methods to improve the processing ability of visual information. Although a great deal of achievements has been made in the modelling of biological visual perception, some of the visual perception models lack the description of biological characteristics, while some are too complex and fail to fully demonstrate the advantages of biological visual information processing².

Until now, the neurophysiological mechanism of human brain memory is still poorly understood, and the research on memory mainly comes from cognitive psychology. In order to simulate the process of memory in the human brain, researchers have carried out a large number of memory experiments and proposed many logical and computational memory models. However, at present, most memory models take vocabulary lists as research objects, and the characteristic expression of memory information is relatively simple, while the memory of visual information is rarely studied³. In recent years, on the basis of cognitive neuroscience research, some neural network models have been established to simulate how the cerebral cortex performs complex and necessary memory functions. However, most of these models are conceptual or abstract models, which are quite complicated to realize and cannot meet the actual processing requirements of natural images.

Therefore, on the basis of anatomical and cognitive neuroscience research, it is of great significance to study how visual information is represented, stored, and remembered in the brain and carry out cognitive neural modelling to realize reliable image classification and recognition. In this book, the visual pathway is modelled based on the mechanism of biological visual perception, and a method of constant feature extraction and expression of biological visual excitation is established. Then, based on the research results of anatomy and cognitive neuroscience, a visual memory model was established to simulate the process of the storage and extraction of visual information by the human brain, combining the memory mechanism of the brain and relevant theories of memory model.

The significance of this book is to break through the traditional visual information processing mode, combine the visual perception and memory mechanism of the human brain with computer vision and image processing technology, and provide a new way for image understanding, recognition and classification.

1.2 Research status of the subject

1.2.1 Review of biological visual perception

The rapid development of brain science and cognitive technology and research methods makes it possible to study the neural mechanism of the structure and function of the biological visual system¹. To explore the neural mechanism of the biological visual perception system and simulate the biological visual perception mechanism to improve the processing ability of computers, visual information processing has become a hot research topic at present.

The human visual system is a powerful system capable of distinguishing millions of different objects⁴. A comprehensive and accurate understanding of the structure and function of the visual system is necessary to build a computational model based on biological vision. Visual information is processed in a hierarchical way in the visual system and transmitted between cortical regions of the brain in accordance with a certain path. The path of visual information transmission is called visual pathway. Ungerleider et al.^1,5,6 believed that the visual system consists of two parallel visual pathways: the ventral pathway and the dorsal pathway, as shown in Fig. 1.1.

Figure 1.1 Two pathways of visual system.

The ventral pathway starts from the retina, passes through the lateral geniculate nucleus (LGN), the visual cortex (V1, V2, V4) areas, the inferior temporal cortex (IT), and finally reaches the ventral prefrontal cortex (VLPFC). This pathway, also known as the What pathway, is responsible for processing static visual information and realizing object perception and recognition. The dorsal pathway starts from the retina, passes through LGN, visual cortex region (V1, V2), middle temporal region (MT), posterior parietal cortex (PPC), and finally reaches the dorsolateral prefrontal cortex (DLPFC). This pathway is also known as the Where pathway, which is responsible for processing dynamic visual information and realizing the recognition of spatial position and movement⁷. The pathway theory of the biological visual system is supported by a large number of anatomical and neurophysiological experimental data, which provides a theoretical basis and research basis for image recognition and classification research based on the mechanism of biological visual perception.

In the real world, there may be large differences between images of the same kind of objects, which are caused by the changes of shape and size within the class, as well as the interference of the shooting Angle, lighting conditions, and the environment. One of the main problems solved by biological vision system is to establish a representation of visual information, which can make object recognition free from interference of factors such as size, contrast, position, angle of view, and light. In order to ensure the accuracy of object recognition, the visual system must not be affected by intra-class differences and be sensitive to small differences between classes, that is, it has good intra-class invariability and inter-class specificity¹. Therefore, how to simulate the powerful biological vision system to extract the invariant features of describing objects has become a key problem to be solved in image recognition and classification.

The ventral pathway is an important pathway for object recognition. The results of cognitive neuroscience research show that the ventral visual cortex usually organizes and processes information in a hierarchical manner, and these structural characteristics play a very important role in the establishment of biological visual perception model. In recent years, based on the relevant research achievements of neurophysiology and cognitive science, scholars at home and abroad have proposed many hierarchical computing models based on biological inspiration to simulate the basic principle and information processing process of the ventral visual pathway.

Neocognitron⁸, proposed by Fukushima, was an early computational model of biological visual perception. Neocognitron is a hierarchical neural network model, whose structure consists of alternating simple neuron layer and complex neuron layer. Simple neuron performs convolution operation, while complex neuron is responsible for pooling and subsampling. The features extracted by this model have certain invariability and selectivity.

The more influential model of biological visual perception is the hierarchical maximization model, HMAX⁹, which was first proposed by Riesenhuber and Poggio in 1999. The model is composed of a linear and nonlinear hierarchical structure, which is used to simulate the hierarchical processing mechanism of visual information in the ventral pathway and explain the information processing in the higher regions of the visual cortex of the brain. In 2007, Serre et al.¹⁰ extended the HMAX model, added the learning p...

Table of contents

• Cover
• Half Title
• Title Page
• Copyright Page
• Contents
• CHAPTER 1: Introduction
• CHAPTER 2: Methods of visual perception and memory modelling
• CHAPTER 3: Bio-inspired model for object recognition based on histogram of oriented gradients
• CHAPTER 4: Modelling object recognition in visual cortex using multiple firing K-means and non-negative sparse coding
• CHAPTER 5: Biological modelling of the human visual system using GLoP filters and sparse coding on multi-manifolds
• CHAPTER 6: Increment learning and rapid retrieval of visual information based on pattern association memory
• CHAPTER 7: Memory modelling based on free energy theory and the restricted Boltzmann machine
• CHAPTER 8: Research on insect pest image detection and recognition based on bio-inspired methods
• CHAPTER 9: Carrot defect detection and grading based on computer vision and deep learning
• CONCLUSIONS