HIERARCHICAL NEURAL NETWORKS FOR IMAGE INTERPRETATION

By N. LUONA

TABLE OF CONTENTS 1 INTRODUCTION MOTIVATION IMPORTANCE OF VISUAL PERCEPTION PERFORMANCE OF THE HUMAN VISUAL SYSTEM LIMITATIONS OF CURRENT COMPUTER VISION SYSTEMS ITERATIVE INTERPRETATION – LOCAL INTERACTIONS IN A HIERARCHY ORGANIZATION OF THE THESIS CONTRIBUTIONS PART I THEORY 2 NEUROBIOLOGICAL BACKGROUND VISUAL PATHWAYS FEATURE MAPS LAYERS NEURONS SYNAPSES DISCUSSION CONCLUSIONS 3 RELATED WORK HIERARCHICAL IMAGE MODELS GENERIC SIGNAL DECOMPOSITIONS NEURAL NETWORKS GENERATIVE STATISTICAL MODELS RECURRENT MODELS MODELS WITH LATERAL INTERACTIONS MODELS WITH VERTICAL FEEDBACK MODELS WITH LATERAL AND VERTICAL FEEDBACK 4 NEURAL ABSTRACTION PYRAMID ARCHITECTURE OVERVIEW HIERARCHICAL NETWORK STRUCTURE DISTRIBUTED REPRESENTATIONS LOCAL RECURRENT CONNECTIVITY ITERATIVE RE?NEMENT FORMAL DESCRIPTION SIMPLE PROCESSING ELEMENTS SHARED WEIGHTS DISCRETE-TIME COMPUTATION VARIOUS TRANSFER FUNCTIONS EXAMPLE NETWORKS LOCAL CONTRAST NORMALIZATION BINARIZATION OF HANDWRITING ACTIVITY-DRIVEN UPDATE INVARIANT FEATURE EXTRACTION 5 UNSUPERVISED LEARNING INTRODUCTION LEARNING A HIERARCHY OF SPARSE FEATURES NETWORK ARCHITECTURE INITIALIZATION HEBBIAN WEIGHT UPDATE COMPETITION LEARNING HIERARCHICAL DIGIT FEATURES DIGIT CLASSI?CATION 6 SUPERVISED LEARNING NEAREST NEIGHBOR CLASSI?ER DECISION TREES BAYESIAN CLASSI?ER SUPPORT VECTOR MACHINES BIAS/VARIANCE DILEMMA FEED-FORWARD NEURAL NETWORKS ERROR BACKPROPAGATION IMPROVEMENTS TO BACKPROPAGATION REGULARIZATION RECURRENT NEURAL NETWORKS BACKPROPAGATION THROUGH TIME REAL-TIME RECURRENT LEARNING DIF?CULTY OF LEARNING LONG-TERM DEPENDENCIES RANDOM RECURRENT NETWORKS WITH FADING MEMORIES ROBUST GRADIENT DESCENT PART II APPLICATIONS 7 RECOGNITION OF METER VALUES INTRODUCTION TO METER VALUE RECOGNITION SWEDISH POST DATABASE PREPROCESSING FILTERING NORMALIZATION BLOCK CLASSI?CATION NETWORK ARCHITECTURE AND TRAINING EXPERIMENTAL RESULTS DIGIT RECOGNITION DIGIT PREPROCESSING COMBINATION WITH BLOCK RECOGNITION 8 BINARIZATION OF MATRIX CODES INTRODUCTION TO TWO-DIMENSIONAL CODES CANADA POST DATABASE ADAPTIVE THRESHOLD BINARIZATION IMAGE DEGRADATION LEARNING BINARIZATION 9 LEARNING ITERATIVE IMAGE RECONSTRUCTION INTRODUCTION TO IMAGE RECONSTRUCTION SUPER-RESOLUTION NIST DIGITS DATASET ARCHITECTURE FOR SUPER-RESOLUTION FILLING-IN OCCLUSIONS MNIST DATASET ARCHITECTURE FOR FILLING-IN OF OCCLUSIONS NOISE REMOVAL AND CONTRAST ENHANCEMENT RECONSTRUCTION FROM A SEQUENCE OF DEGRADED DIGITS 10 FACE LOCALIZATION 101 INTRODUCTION TO FACE LOCALIZATION 102 FACE DATABASE AND PREPROCESSING 103 NETWORK ARCHITECTURE 104 EXPERIMENTAL RESULTS 105 CONCLUSIONS 11 SUMMARY AND CONCLUSIONS 111 SHORT SUMMARY OF CONTRIBUTIONS 112 CONCLUSIONS 113 FUTURE WORK 1131 IMPLEMENTATION OPTIONS 1132 USING MORE COMPLEX PROCESSING ELEMENTS 1133 INTEGRATION INTO COMPLETE SYSTEMS