Intrinsic Decomposition Method Combining Deep Convolutional Neural Network and Probability Graph Model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Current image decomposition methods lack accuracy and recognition rates.
• Increasing demand for high-quality image decomposition in AI and computer vision.

Purpose of the Study:

• To develop a highly accurate single-image intrinsic image decomposition method.
• To improve upon existing automatic decomposition algorithms using CNNs and probability maps.

Main Methods:

• Combined convolutional neural network (CNN) with a probability map model.
• Developed a single-image intrinsic image decomposition algorithm.
• Proposed a multi-image collaborative intrinsic image decomposition method.
• Applied eigenimage decomposition to illumination uniformity in change detection.

Main Results:

• Achieved visual effects comparable to user-interactive decomposition.
• Obtained the lowest error rate on standard dataset images compared to existing algorithms.
• Demonstrated consistent foreground reflectivity in multi-image decomposition.
• Improved accuracy in cooperative saliency detection using reflectivity layer images.

Conclusions:

• The proposed CNN and probability graph model cooperation enhances pixel-level eigendecomposition.
• The method shows superior performance on the Msrc-v2 dataset, with a 0.8% improvement over probability plot models.