دانلود کتاب Pattern Recognition and Computer Vision: 4th Chinese Conference, PRCV 2021, Beijing, China, October 29 – November 1, 2021, Proceedings, Part IV (Image ... Vision, Pattern Recognition, and Graphics)

فهرست مطالب :

Preface
Organization
Contents - Part IV
Machine Learning, Neural Network and Deep Learning
Edge-Wise One-Level Global Pruning on NAS Generated Networks
1 Introduction
2 Related Work
2.1 Neural Architecture Search
2.2 Network Pruning
3 Edge-Wise One-Level Global Pruning on DARTS-Based Network
3.1 Edge Weight Assignment
3.2 One-Level Structure Learning
3.3 Procedure of EOG-Pruning
3.4 Comparison with Previous Work
4 Experiments
4.1 Datasets
4.2 EOG-Pruning on CIFAR-10 and CIFAR-100
4.3 EOG-Pruning on ImageNet
4.4 Visualization
4.5 Ablation Study
5 Conclusion
References
Convolution Tells Where to Look
1 Introduction
2 Related Work
3 Feature Difference Module
3.1 Feature Difference
3.2 FD Networks
4 Experiments
4.1 Results on CIFAR-10 and CIFAR-100
4.2 ImageNet Classification
4.3 VOC 2012 Object Detection
5 Analysis and Interpretation
6 Conclusion
References
Robust Single-Step Adversarial Training with Regularizer
1 Introduction
2 Related Work
2.1 Adversarial Training
2.2 Single-Step Adversarial Training
3 Proposed Approach
3.1 Observation
3.2 PGD Regularization
3.3 Training Route
4 Experiments
4.1 Results on MNIST
4.2 Results on CIFAR-10
5 Conclusion
References
Texture-Guided U-Net for OCT-to-OCTA Generation
1 Introduction
2 Method
3 Experiments
3.1 Dataset and Metrics
3.2 Results
4 Conclusion
References
Learning Key Actors and Their Interactions for Group Activity Recognition
1 Introduction
2 Related Work
2.1 Group Activity Recognition
2.2 Graph Neural Networks
3 Approach and Framework
3.1 Preliminaries
3.2 Extracting the SARG
3.3 Feature Fusion
3.4 Training Loss
4 Experiments
4.1 Ablation Studies
4.2 Compared with SOTA
5 Conclusion
References
Attributed Non-negative Matrix Multi-factorization for Data Representation
1 Introduction
2 Related Work
2.1 NMF
2.2 GNMF
3 The Proposed Method
3.1 Motivation and Objective Function
3.2 Model Optimization
3.3 Algorithm Analysis
4 Experiments
4.1 Experiment Settings
4.2 Experiment Results
4.3 Parameter Analysis
4.4 Convergence Study
5 Conclusions
References
Improved Categorical Cross-Entropy Loss for Training Deep Neural Networks with Noisy Labels
1 Introduction
2 Improved Categorical Cross-Entropy Loss for Noise-Robust Classification
2.1 Robustness Analysis of CCE and MAE
2.2 Improved Categorical Cross Entropy
2.3 Theoretical Analysis of ICCE
3 Experiment
3.1 Dataset and Model Architectures
3.2 Label Noise
3.3 Evaluation of the CIFAR Dataset with Synthetic Noise
4 Conclusion
References
A Residual Correction Approach for Semi-supervised Semantic Segmentation
1 Introduction
2 Related Work
2.1 Supervised Semantic Segmentation
2.2 Semi-supervised Semantic Segmentation
3 Method
3.1 Residual Correction Network
3.2 Supervised Training
3.3 Semi-supervised Training
4 Experiments
4.1 Network Architecture
4.2 Datasets and Evaluation Metrics
4.3 Experimental Settings
4.4 Results
5 Conclusions
References
Hypergraph Convolutional Network with Hybrid Higher-Order Neighbors
1 Introduction
2 Related Works
2.1 Graph Neural Networks
2.2 Learning on Hypergraph
2.3 Hypergraph Neural Network
3 Preliminary Knowledge
3.1 Hypergraph
3.2 Hypergraph Convolution Network
4 Method
5 Experiments
5.1 Datasets and Baseline
5.2 Experimental Setting
5.3 Experimental Results and Discussion
6 Conclusions
References
Text-Aware Single Image Specular Highlight Removal
1 Introduction
2 Related Work
2.1 Dichromatic Reflection Model-Based Methods
2.2 Inpainting-Based Methods
2.3 Deep Learning-Based Methods
3 Datasets
3.1 Real Dataset
3.2 Synthetic Datasets
4 Proposed Method
4.1 Network Architecture
4.2 Loss Functions
5 Experiments
5.1 Implementation Settings
5.2 Qualitative Evaluation
5.3 Quantitative Evaluation
5.4 Ablation Study
6 Conclusion and Future Work
References
Minimizing Wasserstein-1 Distance by Quantile Regression for GANs Model
1 Introduction
2 Related Work
3 Background
3.1 Wasserstein GAN
3.2 Quantile Regression
4 Our Method
5 Experiment
5.1 Basic Setting
5.2 IS and FID Criteria Experiment
5.3 Adam and Rmsprob Experiment
5.4 Different Quantile Initialization Experiment
6 Conclusion
References
A Competition of Shape and Texture Bias by Multi-view Image Representation
1 Introduction
2 Multi-view Image Representations
2.1 Background Representations
2.2 Shape Representations
2.3 Texture Representations
3 Loss
4 Experiments
4.1 Datasets
4.2 Implementation Details
4.3 Results and Analysis
5 Conclusion and Future Work
References
Learning Indistinguishable and Transferable Adversarial Examples
1 Introduction
2 Related Work
2.1 Adversarial Example Generation
2.2 The Fast Gradient Sign Method Series
3 Methodology
3.1 The Piecewise Linear Function
3.2 The Gradient Regularization Term
3.3 The Mixed-Input Strategy
4 Experiments
4.1 Experimental Setting
4.2 Ablation Study
4.3 The Results of Single-Model Attacks
4.4 The Results of Multi-model Attacks
5 Conclusion
References
Efficient Object Detection and Classification of Ground Objects from Thermal Infrared Remote Sensing Image Based on Deep Learning
1 Introduction
2 Related Work
3 Methodology
3.1 Backbone Network
3.2 Feature Pyramid Networks
3.3 Adaptive Multiscale Receptive Field
3.4 Loss Function
4 Experiments and Results
4.1 Dataset
4.2 Results
5 Conclusions
References
MEMA-NAS: Memory-Efficient Multi-Agent Neural Architecture Search
1 Introduction
2 MEMA-NAS Framework
2.1 Multi-Agent Search Algorithm
2.2 Resource Constraint
3 Experiments
3.1 Datasets and Evaluations
3.2 Illustration of Searched Architecture and Analysis
3.3 Ablation Experiments for Individual Module Search
3.4 Generalization Performance Evaluation
3.5 Comparison with NAS-Based Detection Works
3.6 GPU Memory Consumption Evaluation
4 Conclusion
References
Adversarial Decoupling for Weakly Supervised Semantic Segmentation
1 Introduction
2 Related Work
2.1 Initial Prediction for WSSS
2.2 Response Refinement for WSSS
3 Method
3.1 Sub-category Clustering
3.2 Context Decoupling Augmentation
3.3 Adversarial Climbing
4 Experiments
4.1 Evaluated Dataset and Metric
4.2 Ablation Study and Analysis
4.3 Semantic Segmentation Performance
5 Conclusion
References
Towards End-to-End Embroidery Style Generation: A Paired Dataset and Benchmark
1 Introduction
2 Related Work
2.1 Traditional Stitching Generation
2.2 Style Transfer
2.3 Image-to-Image Translation
2.4 Related Datasets
3 Paired Embroidery Dataset Overview
4 Automatic Embroidery Generation
4.1 Local Texture Generation
4.2 Global Fine-Tuning
5 Experiments
5.1 Experimental Setting
5.2 Evaluation Metrics
5.3 Experimental Results
6 Conclusion
References
Efficient and Real-Time Particle Detection via Encoder-Decoder Network
1 Introduction
2 Related Work
2.1 Particle Detection Methods
2.2 Decoder Architectures
2.3 Light-Weight Networks
2.4 Knowledge Distillation
3 The Proposed Method
3.1 Network Framework
3.2 Loss Function
3.3 Structured Knowledge Distillation
4 Experiments
4.1 Datasets and Implementation Details
4.2 Evaluation Metrics
4.3 Results and Analysis
5 Conclusion
References
Flexible Projection Search Using Optimal Re-weighted Adjacency for Unsupervised Manifold Learning
1 Introduction
2 Graph Construction and Metric Learning
2.1 Graph-Based Affinity Matrix
2.2 Graph Embedding
2.3 Distance Metric Learning
3 Flexible Projection Search Using Optimal Re-weighted Adjacency (FPSORA)
3.1 Quadratic Programming for Updating Adjacency Weights
3.2 Flexible Projection Search for Preservation of Original Locality
3.3 Complete Workflow of the Proposed FPSORA
4 Experiments and Analysis
4.1 Evaluation Metrics
4.2 Experimental Setup
4.3 Experiments and Analysis
5 Conclusion
References
Fabric Defect Detection via Multi-scale Feature Fusion-Based Saliency
1 Introduction
2 Proposed Method
2.1 Multi-scale Feature Learning Module
2.2 Feedback Attention Refinement Fusion Module
2.3 The Joint Loss
3 Experiments
3.1 Dataset and Evaluation Metrics
3.2 Implementation Details
3.3 Comparison with State-of-the-Arts
3.4 Ablation Study
4 Conclusion
References
Improving Adversarial Robustness of Detector via Objectness Regularization
1 Introduction
2 Method
2.1 Vanishing Adversarial Patch
2.2 Objectness Regularization
3 Experiment
3.1 Experimental Setup
3.2 Experimental Result
4 Conclusion
References
IPE Transformer for Depth Completion with Input-Aware Positional Embeddings
1 Introduction
2 Related Work
3 Method
3.1 Input-Aware Positional Embedding
3.2 IPE Transformer
3.3 Network Structure
4 Experiment
4.1 Setup
4.2 Experiment Results
4.3 Ablation Study
5 Conclusion
References
Enhanced Multi-view Matrix Factorization with Shared Representation
1 Introduction
2 Related Work
3 Proposed Method
3.1 Multi-view Matrix Factorization with Shared Representation
3.2 Enhancement of Feature Extraction Capability
4 Experiments
4.1 Datasets
4.2 Experimental Setup
4.3 Experimental Results
4.4 Ablation Study
5 Conclusion
References
Multi-level Residual Attention Network for Speckle Suppression
1 Introduction
2 Related Work
3 Proposed Method
4 Experimental Results
4.1 Datasets and Evaluation Index
4.2 Experiment on Synthetic Speckled Image
4.3 Experiment on Real Image
5 Conclusion
References
Suppressing Style-Sensitive Features via Randomly Erasing for Domain Generalizable Semantic Segmentation
1 Introduction
2 Related Work
2.1 Semantic Segmentation
2.2 Domain Generalization
3 Method
3.1 Overview
3.2 Framework
3.3 Randomly Erasing Style-sensitive-Channel Features
3.4 Training Loss
4 Experiment
4.1 Datasets
4.2 Implementation Details
4.3 Performance Comparison
4.4 Ablation Study
5 Conclusion
References
MAGAN: Multi-attention Generative Adversarial Networks for Text-to-Image Generation
1 Introduction
2 Related Work
2.1 Text-to-Image Generation
2.2 Attention Mechanism
2.3 Multi-head Mechanism
3 Multi-Attention Gan (MAGAN)
3.1 Model Overview
3.2 Self-attention Module
3.3 Multi-head Attention Module
3.4 Objective Functions
4 Experiments
4.1 Datasets
4.2 Evaluation Metric
4.3 Implementation Details
4.4 Quantitative Results
4.5 Qualitative Results
4.6 Ablation Studies
5 Conclusion
References
Dual Attention Based Network with Hierarchical ConvLSTM for Video Object Segmentation
1 Introduction
2 Related Work
2.1 Semi-supervised Video Object Segmentation
3 Method
3.1 Framework
3.2 Encoder
3.3 Recurrent ConvLSTM Decoder
3.4 Loss Function
4 Experiments
4.1 Implementation Details
4.2 Dataset
4.3 Ablation Study
4.4 Semi-supervised Video Object Segmentation
4.5 Visualization
5 Conclusion
References
Distance-Based Class Activation Map for Metric Learning
1 Introduction
2 Distance-Based Class Activation Map
3 Experiments
4 Metric-Based Applications
4.1 Few-Shot Learning
4.2 Image Retrieval
4.3 Re-identification
5 Conclusion
References
Reading Pointer Meter Through One Stage End-to-End Deep Regression
1 Introduction
2 Related Work
2.1 Recognition by Dial Detection
2.2 Recognition by Keypoint Detection
2.3 Direct Recognition Through Deep Regression
3 Method Proposed in the Paper
3.1 Backbone Network
3.2 Loss Function
3.3 Model Training
4 Experimental Results
4.1 Dataset
4.2 Experiments
4.3 Additional Experiments
5 Conclusion
References
Deep Architecture Compression with Automatic Clustering of Similar Neurons
1 Introduction
2 Related Work
2.1 Pruning
2.2 Knowledge Distillation
2.3 Lightweight Network Architecture Design
3 Automatic Neurons Clustering
3.1 Neurons Redundancy Analysis
3.2 Automatic Neurons Clustering Algorithm
4 Experiment and Analysis
4.1 Experiment on the Fully Connected Network
4.2 Experiment on Convolutional Networks
5 Conclusion and Discussion
References
Attention Guided Spatio-Temporal Artifacts Extraction for Deepfake Detection
1 Introduction
2 Related Work
3 The Proposed Method
3.1 Face Feature Extraction Module
3.2 Attention Guided LSTM Module
3.3 Classfication
4 Experiments
4.1 Datasets
4.2 Setup and Implementation Details
4.3 Comparison with State-of-the-Art Methods
4.4 Ablation Study
4.5 Visualization
5 Conclusion
References
Learn the Approximation Distribution of Sparse Coding with Mixture Sparsity Network
1 Introduction
2 The Proposed Method
2.1 LISTA Network
2.2 Mixture Sparsity Network
3 Experiment
3.1 Synthetic Experiments
3.2 Digit Data Experiments
3.3 Sparsity Analysis
4 Conclusion
References
Anti-occluded Person Re-identification via Pose Restoration and Dual Channel Feature Distance Measurement
1 Introduction
2 Related Works
2.1 Traditional ReID
2.2 Occluded ReID
3 Proposed Method
3.1 Person Pose Repair
3.2 Dual Channel Feature Extraction
4 Experiment
4.1 Datasets and Evaluation Metrics
4.2 Implementation Details
4.3 Comparative Experiments
4.4 Visualization
4.5 Ablation Study
5 Conclusion
References
Dynamic Runtime Feature Map Pruning
1 Introduction
2 Runtime Feature Map Pruning
2.1 Dynamic Pruning
2.2 Dynamic Pruning Algorithm
2.3 Bandwidth Reduction in Video Streams
3 Experiments
4 Related Work
5 Future Work
References
Special Session: New Advances in Visual Perception and Understanding
Multi-branch Graph Network for Learning Human-Object Interaction
1 Introduction
2 Related Work
3 Approach
3.1 Overview
3.2 Object Embedding
3.3 Multi-branch Network
3.4 Graph Network
4 Experiments
4.1 Dataset and Evaluate Metric
4.2 Implementation Details
4.3 Comparison with State-of-the-Art Methods
4.4 Ablation Study
5 Conclusion
References
FDEA: Face Dataset with Ethnicity Attribute
1 Introduction
2 Proposed FDEA
2.1 Images Extracted from CelebA and Annotation
2.2 Image Collection from Annotated Datasets
2.3 Images Crawled from the Internet
3 Experimental Results
4 Conclusion
References
TMD-FS: Improving Few-Shot Object Detection with Transformer Multi-modal Directing
1 Introduction
2 Related Work
3 Methodology
3.1 FSOD Fine-Tuning
3.2 Semantic and Visual Alignment (SVA)
3.3 Semantic Visual Transformer (SVT)
3.4 Loss
4 Experiments
4.1 Implementation Details
4.2 Existing Benchmarks
4.3 Ablation Study
4.4 Visualization
5 Conclusion
References
Feature Matching Network for Weakly-Supervised Temporal Action Localization
1 Introduction
2 Related Work
3 Network Structure
3.1 Problem Statement
3.2 Network Structure
3.3 Attention Branch
3.4 Temporal Action Localization
4 Experiments
4.1 Datasets and Evaluation Metrics
4.2 Implementation Details
4.3 Ablation Studies
5 Conclusion
References
LiDAR-Based Symmetrical Guidance for 3D Object Detection
1 Introduction
2 Related Work
3 Method
3.1 Generation of Symmetric Annotations
3.2 SS-PV-RCNN
3.3 Consistency Loss
4 Experiments
4.1 Symmetry Enhances Data Analysis
4.2 Comparison with State-of-the-Arts
4.3 Ablation Study
5 Conclusion
References
Few-Shot Segmentation via Complementary Prototype Learning and Cascaded Refinement
1 Introduction
2 Method
2.1 Method Overview
2.2 Complementary Prototype Learning and Cascaded Refinement
2.3 Loss Function
3 Experiments
3.1 Datasets and Evaluation Metric
3.2 Implementation Details
3.3 Comparison to State-of-the-Art
3.4 Ablation Study
4 Conclusion
References
Couple Double-Stage FPNs with Single Pipe-Line for Solar Speckle Images Deblurring
1 Introduction
2 Method Description
2.1 Main Idea
2.2 Description of the Generator
2.3 Discriminators
3 Experiments and Analysis
3.1 Solar Speckle Datasets
3.2 Implementation Details
3.3 Experimental Evaluation
3.4 Ablation Experiment
3.5 Comparison Results on Public Datasets
4 Conclusion
References
Multi-scale Image Partitioning and Saliency Detection for Single Image Blind Deblurring
1 Introduction
2 Related Work
3 Multi-scale Adaptive Image Partition and Merging
4 Blind Deblurring Based on Saliency Detection
4.1 Saliency Detection Based on SUN
4.2 Blind Deblurring Model and Solution
5 Weighted Window Function for Image Patches Joint with Different Sizes
6 Experiments and Results
6.1 The Comparisons with the State-of-the-Art Methods on the Space-Invariant Blurred Image
6.2 Quantitative Evaluations
6.3 Comparison with the State-of-the-Art Methods on the Space-Variant Blurred Image
7 Conclusion
References
CETransformer: Casual Effect Estimation via Transformer Based Representation Learning
1 Introduction
2 Preliminary and Background
3 Methodology
3.1 Overview of the Proposed CETransformer Model
3.2 Self-supervised Transformer
3.3 Adversarial Learning for Distribution Balancing
3.4 Outcome Prediction
4 Experiments
4.1 Datasets and Metric
4.2 Competing Algorithms
4.3 Prediction Performance Results
4.4 Ablation Study
5 Conclusion
References
An Efficient Polyp Detection Framework with Suspicious Targets Assisted Training
1 Introduction
2 Methods
2.1 Multi-branch Spatial Attention Mechanism
2.2 Top Likelihood and Similarity Loss
2.3 Cross Stage Partial Connection
3 Experiment
3.1 Datasets
3.2 Evaluation and Results
4 Conclusions
References
Invertible Image Compressive Sensing
1 Introduction
2 Related Work
2.1 Compressive Sensing
2.2 Invertible Neural Network
3 Proposed InvICS
3.1 Sampling Subnet
3.2 Initialization Subnet
3.3 Invertible Recovery Subnet
3.4 Network Parameters and Loss Function
4 Analysis and Experiments
4.1 Study of Phase Number
4.2 Comparison with State-of-the-Art
4.3 Ablation Studies
5 Conclusion and Future Work
References
Gradient-Free Neural Network Training Based on Deep Dictionary Learning with the Log Regularizer
1 Introduction
2 Related Work
2.1 Notation
2.2 Neural Network Training and Deep Dictionary Learning
3 Gradient-Free Neural Network Training Based on Deep Dictionary Learning with the Log Regularizer
3.1 Problem Formulation
3.2 Proposed Algorithm
3.3 The Overall Algorithm
4 Numerical Experiments
4.1 Parameter Setting
4.2 Classification Experiments on MNIST
4.3 Classification Experiments on Fashion MNIST
5 Conclusion and Future Work
References
Author Index