Unsupervised Intrinsic Image Decomposition Using Internal Self-Similarity Cues.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Supervised learning methods for intrinsic image decomposition require extensive ground truth data, limiting their use on real-world images.
• Acquiring ground truth intrinsic decomposition for natural images is a significant challenge.

Purpose of the Study:

• To develop an unsupervised framework for intrinsic image decomposition from single natural images.
• To overcome the limitations of supervised methods by eliminating the need for ground truth data.

Main Methods:

• An unsupervised intrinsic decomposition network (UIDNet) with two sub-networks: reflectance prediction network (RPN) and shading prediction network (SPN).
• The network leverages the internal self-similarity of reflectance patches and convolutional network properties.
• A novel loss function is designed to promote reflectance self-similarity and jointly train the RPN and SPN to reconstruct the input image.

Main Results:

• The proposed unsupervised framework effectively learns intrinsic image decomposition from single images.
• Experimental results on three benchmark datasets demonstrate the superiority of the UIDNet.
• The method successfully decomposes images into reflectance and shading components without ground truth.

Conclusions:

• Unsupervised intrinsic image decomposition is feasible and effective using the proposed UIDNet.
• The approach offers a practical solution for decomposing real-world natural images.
• Promoting reflectance self-similarity is a key factor in successful unsupervised decomposition.