Painting Element Segmentation Algorithm Based on Deep Network
Song Ran, Lan Lin · 2021
We know that art and technology complement each other, because the art puts forward new requirements and development directions for technology, and technology supports the art performance more colorful. In recent years, the in-depth study of deep network has greatly promoted the progress of image processing technology, especially the combination of deep network and painting art has become a hot topic. Segmentation is of great importance in image processing. Painting elements segmentation and extraction is the basis of further research, such as identification, comparison and restoration in the later stage. There are many defects and shortcomings in traditional image segmentation such as low accuracy, noise sensitivity. While the method based on deep network shows many advantages, which can solve the problem of noise and non-uniformity in the image.In this paper, U-Net and IndexNet are selected to study the Figure segmentation of ancient painting. There are a few ancient painting samples for training network. If the network is trained directly, it cannot overcome the over-fit problem. A data enhancement method is proposed to improve the data volume and the reliability, then the data is used for U-Net training, and the loss function is minimized. There are four subsampling operations in the whole U-Net network. The operation of the up-sampling part is the opposite of the down-sampling, with 4 up-sampling times. The self-made dataset is put into the corresponding folder, and the appropriate learning rate is chosen through experiment. IndexNet is also trained and parameterized to produce predictions. Then the differences between the two models are analyzed and compared, and the better prediction results are selected for guided filtering. The guided filter is used for image dehazing and extinction, and when the guided image is the original image, the filter becomes an edge-preserving filter to optimize the prediction results. In the segmentation task of ancient paintings, we combine the segmentation network with the guided filtering network. There is high segmentation accuracy with our method. The Figure achieves a complete separation from the background.