This paper presents a neural network approach for image completion that maintains both global and local consistency. The method uses a fully-convolutional neural network to complete images of arbitrary resolutions with missing regions of any shape. To ensure coherence, the system employs two discriminator networks — a global discriminator that evaluates overall image consistency and a local discriminator that assesses the quality of small regions around completed areas. Unlike patch-based methods, this approach can synthesize novel image fragments not present elsewhere in the source image, enabling natural completion of structured objects.

Image completion (or inpainting) aims to fill missing regions in a way that is visually plausible. Traditional patch-based methods like PatchMatch search for similar patches within the image to fill holes. While effective for texture synthesis and repetitive patterns, they cannot generate semantically meaningful content absent from the image — for example, completing a face when the eye region is missing. Deep learning approaches like Context Encoders demonstrated that neural networks can learn to hallucinate plausible content, but results suffered from blurriness and lacked local detail consistency.

The key insight of this work is that ensuring both global structure and local texture quality requires separate evaluation mechanisms. A single discriminator struggles to balance these two aspects. The authors propose a completion network trained with two adversarial discriminators: one sees the entire image to judge global coherence (e.g., scene layout, object proportions), and the other focuses on a small region centred on the completed area to judge local texture fidelity.

Diffusion-based methods propagate information from boundaries inward, working well for narrow missing regions but failing for large holes. Patch-based methods (PatchMatch and variants) iteratively search and paste the best matching patches, excelling at texture propagation but lacking semantic understanding. Deep generative models, particularly Generative Adversarial Networks (GANs), have opened new possibilities by learning to generate realistic image content. Context Encoders combined an encoder-decoder architecture with adversarial training for inpainting, while semantic inpainting methods optimise in the latent space of pre-trained generators. These approaches typically use only a single discriminator and produce results with noticeable artifacts at completion boundaries.

The completion network is a fully convolutional architecture consisting of downsampling layers, dilated convolutional layers, and upsampling layers. Dilated convolutions are used in the bottleneck to increase the receptive field without reducing spatial resolution, allowing the network to aggregate information from a larger context when filling holes. The network accepts an input image with a binary mask indicating missing regions and outputs the completed image. It is fully convolutional, meaning it can handle arbitrary input resolutions and arbitrarily shaped missing regions at test time.

The training framework employs two adversarial discriminator networks. The global discriminator receives the entire completed image as input and evaluates whether the overall scene is coherent — checking for consistent lighting, perspective, and semantic layout. The local discriminator receives a small crop centred on the completed region and evaluates whether the local texture and structure are realistic. The total adversarial loss is a weighted combination of both discriminators' outputs. These discriminators serve as auxiliary training networks — they guide the completion network during training but are discarded at inference time.

The method is evaluated on Places2 for scene completion and CelebA for face inpainting. Compared to PatchMatch and Context Encoders, the proposed method produces significantly more natural completions, especially for large missing regions and structured content like faces. The global discriminator prevents inconsistencies in scene layout, while the local discriminator improves texture sharpness. Ablation studies demonstrate that removing either discriminator degrades quality: without the global discriminator, completions have good local texture but inconsistent global structure; without the local discriminator, results are globally coherent but locally blurry. The method handles object removal, scene completion, and arbitrary-shaped holes.

This paper proposes a deep image completion method that ensures both global and local consistency through a dual-discriminator training framework. The fully-convolutional completion network can process arbitrary resolutions and hole shapes, while the global and local discriminators guide it to produce completions that are both semantically coherent and texturally detailed. The approach significantly outperforms patch-based methods and prior deep learning approaches, particularly for structured content and large missing regions. The dual-discriminator design is general and could benefit other image generation tasks.