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

With the rapid development of computer vision and artificial intelligence, people are increasingly demanding image decomposition. Many of the current methods do not decompose images well. In order to find the decomposition method with high accuracy and accurate recognition rate, this study combines convolutional neural network and probability map model, and proposes a single-image intrinsic image decomposition method that is on both standard dataset images and natural images. Compared with the existing single-image automatic decomposition algorithm, the visual effect comparable to the user interaction decomposition algorithm is obtained, and the method of this study also obtains the lowest error rate in the quantitative comparison on the standard dataset image. The multi-image collaborative intrinsic image decomposition method proposed in this study obtains the decomposition result of consistent foreground reflectivity on multiple sets of image pairs. In this study, the eigenimage decomposition is applied to the illumination uniformity in the small change detection, and the promising reflectivity layer image obtained by the decomposition helps to improve the accuracy of the cooperative saliency detection. This study proposes an algorithm for the cooperation between CNN and probability graph model, and introduces how to combine the probability graph model with the traditional CNN to accomplish the pixel-level eigendecomposition task. This study also designs a single-image and multi-image intrinsic image decomposition results analysis experiments, then analyzes the probabilistic graphical model coordination intrinsic image decomposition results, and finally analyzes the convolutional neural network coordination intrinsic decomposition performance to draw the conclusion of this study. The effect on the Msrc-v2 dataset was increased by 0.8% over the probability plot model.