Phân đoạn hình ảnh ngữ nghĩa trong điều kiện tối với phương pháp thích ứng miền

2021

101
0
0

Phí lưu trữ

30.000 VNĐ

Mục lục chi tiết

Acknowledgement

ABSTRACT

1. CHƯƠNG 1: INTRODUCTION

1.1. Practical Context

1.2. Problem Definition

2. CHƯƠNG 2: RELATED WORK

2.1. Fundamental Knowledge in Convolutional Neural Network

2.2. Semantic Image Segmentation

2.2.1. Overview of Semantic Image Segmentation

2.2.2. Fully Convolutional Networks

2.2.3. Pyramid Scene Parsing Network

2.2.4. Google DeepLab Family

2.3. Generative Adversarial Network

2.3.1. Overview of Generative Adversarial Network

2.3.2. Conditional Generative Adversarial Network

2.3.3. Pix2Pix

2.3.4. CycleGAN

2.3.5. Image Domain Adaptation

3. CHƯƠNG 3: PROPOSED FRAMEWORK

3.1. GAN-based Image Translation Component

3.1.1. Variational Autoencoders-GAN

3.1.2. Perceptual loss maintains the semantic features

3.2. Semantic Image Segmentation with Self-training Strategy

3.2.1. Panoptic Feature Pyramid Networks

3.2.2. Proposed Loss Function

3.2.3. Our Self-training Method

4. CHƯƠNG 4: EXPERIMENTS

4.1. Datasets for Image Domain Translation

4.2. Datasets for Semantic Image Segmentation

4.2.1. Nighttime Driving Testset

4.2.2. Extra Unlabeled Data Selection

4.3. Day2Night Image Domain Translation Component

4.3.1. With Perceptual Loss Refinement Results

4.4. Semantic Image Segmentation Component

4.4.1. Daytime Cityscapes Images Training

4.4.2. Daytime and Nighttime Images Training

4.4.3. Daytime and Nighttime Images Training with Perceptual Loss for Image Translation Module

4.4.4. Only-night Images Training with Perceptual Loss for Image Translation Module

4.4.5. Daytime and Nighttime Images Training with Focal Loss

4.4.6. Daytime and Nighttime Images Training with FID-based Method for Extra Unlabeled Data

4.5. Lessons from Series of Experiments

4.5.1. Improving GAN-based Method

4.5.2. Improving Semantic Image Segmentation Component

Publication

References