Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

EdgeConnect: Generative Image Inpainting with Adversarial Edge Learning (1901.00212v3)

Published 1 Jan 2019 in cs.CV

Abstract: Over the last few years, deep learning techniques have yielded significant improvements in image inpainting. However, many of these techniques fail to reconstruct reasonable structures as they are commonly over-smoothed and/or blurry. This paper develops a new approach for image inpainting that does a better job of reproducing filled regions exhibiting fine details. We propose a two-stage adversarial model EdgeConnect that comprises of an edge generator followed by an image completion network. The edge generator hallucinates edges of the missing region (both regular and irregular) of the image, and the image completion network fills in the missing regions using hallucinated edges as a priori. We evaluate our model end-to-end over the publicly available datasets CelebA, Places2, and Paris StreetView, and show that it outperforms current state-of-the-art techniques quantitatively and qualitatively. Code and models available at: https://github.com/knazeri/edge-connect

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Kamyar Nazeri (6 papers)
  2. Eric Ng (3 papers)
  3. Tony Joseph (3 papers)
  4. Faisal Z. Qureshi (13 papers)
  5. Mehran Ebrahimi (17 papers)
Citations (655)

Summary

EdgeConnect: Generative Image Inpainting with Adversarial Edge Learning

The paper "EdgeConnect: Generative Image Inpainting with Adversarial Edge Learning" presents a novel methodology for addressing the problem of image inpainting. This approach leverages a dual-stage process for inpainting missing regions in images, focusing on generating coherent textures and edges.

Methodology

EdgeConnect introduces a unique two-part architecture involving edge generation and contextual completion. The first part is an edge generator, which is designed to predict edge maps from corrupted images using adversarial training. The second part employs a completion network that uses these predicted edge maps to reconstruct the missing content in the image, thereby enhancing the plausibility and aesthetic quality of the generated regions.

Network Architecture

The network architecture capitalizes on a Generative Adversarial Network (GAN) framework. This includes a generator that crafts realistic-looking edges, countered by a discriminator that assesses the authenticity of these edges. The overall architecture is characterized by its ability to produce sharp and coherent results even when dealing with large missing regions in images.

Experiments and Results

The empirical evaluation of the EdgeConnect method is detailed through extensive experiments on standard datasets, including CelebA-HQ and Places2. The performance metrics compare favorably against several state-of-the-art inpainting methods. Quantitative assessments via SSIM and PSNR indicate superior performance, while qualitative evaluations demonstrate the capability of the model in handling complex image structures and maintaining consistency in textures and semantics.

Discussions

One of the notable advantages of the EdgeConnect framework is its ability to handle complex images with substantial region removal while still preserving contextual integrity. The separation of edge generation from contextual inpainting allows the model to focus on structural elements first, leading to more coherent and visually appealing results.

Implications and Future Work

The implications of this research are notable in fields such as digital restoration, content editing, and automated image correction. The introduction of an edge-based adversarial approach provides a template for future developments in image processing tasks, potentially extending to video inpainting and other high-dimensional data restoration applications.

Future work suggested in this domain includes refinements in edge prediction accuracy and extending the model’s capability to multichannel and high-resolution images without significant performance degradation. Researchers could focus on integrating this method with other styles of data input or further refining the adversarial learning mechanisms to optimize both efficiency and output quality.

In conclusion, the EdgeConnect paper contributes a significant advancement in the field of image inpainting, offering promising directions for future exploration and applications in computer vision.