Mural is an important component of culture and art of Dunhuang in China. Unfortunately, these murals had been ruined or are being ruined by some diseases such as cracking, hollowing, falling off, mildew, dirt, and so on. Existing image restoration algorithms have problems such as incomplete repair and disharmonious texture during large-area repair, so the effect of mural image disease area repair is poor. Due to lack of a standard mural datasets, Dunhuang mural datasets are created in the paper. Meanwhile, our network architecture SeparaFill is proposed which connects two generators based on U-Net. Based on the characteristics of the painting, the contour line pixel area of the mural image is innovatively separated from the content pixel area. Firstly, the contour restoration generator network with skip connect and hierarchical residual blocks is employed to repair contour lines. Then, the color mural image is repaired by the content completion network with guide of the repaired contour. Full resolution branches and generator branches of the U type are exploited in content completion generators. Convolution layers of different kernel sizes are fused to improve the reusability of the underlying features. Finally, global and local discriminant networks are applied to determine whether the repaired mural image is authentic in terms of both the modified and unmodified areas. The proposed SeparaFill shows good performance in restoring the line structure of the damaged mural images and retaining the contour information of the mural images. Compared with existing restoration algorithms in mural real damage repair experiment, our algorithm increases the peak signal-to-noise ratio (PSNR) by an average of 1.1–4.3 dB and the structural similarity (SSIM) values were slightly improved. Experimental results reveal the good performance of the proposed model, which can contribute to digital restorations of ancient murals.

In recent years, with the development of the internet, video has become more and more widely used in life. Adding harmonious music to a video is gradually becoming an artistic task. However, artificially adding music takes a lot of time and effort, so we propose a method to recommend background music for videos. The emotional message of music is rarely taken into account in current work, but it is crucial for video music retrieval. To achieve this, we design two paths to process content information and emotional information between modals. Based on the characteristics of video and music, we design various feature extraction schemes and common representation spaces. In the content path, the pre-trained network is used as the feature extraction network. As these features contain some redundant information, we use an encoder–decoder structure for dimensionality reduction. Where encoder weights are shared to obtain content sharing features for video and music. In the emotion path, an emotion key frames scheme was used for video and a channel attention mechanism was used for music in order to obtain the emotion information effectively. We also added emotion distinguish loss to guarantee that the network acquires the emotion information effectively. More importantly, we propose a way to combine content information with emotional information. That is, content features are first stitched together with sentiment features and then passed through a fused shared space structured as an MLP to obtain more effective fused shared features. In addition, a polarity penalty factor has been added to the classical metric loss function to make it more suitable for this task. Experiments show that this dual path video music retrieval network can effectively merge information. Compared with existing methods, the retrieval task evaluation index increases Recall@1 by 3.94.

To get the closed boundary of target object, this paper proposed an improved algorithm with disparity-guided constraint for active contour extraction based on greedy snake model. The disparity relationship between the control point and the centre of the object could introduce a new external force, which adds the function of shrinking and expanding for the traditional greedy snake when the curve falls onto texture-less regions. In addition, our method also uses the improved model iteratively with processed control points as the original input. Experiment results have proved this method will obtain the accurate contour made up of uniform and dense points, which is more effective than the traditional greedy snake algorithm.

A novel and convenient method for human positioning based on computer vision is proposed. The new camera Realsense D435 from Intel is used to provide richer and more effective visual information. The positioning problem can be simplified to the conversion problem between point cloud and space coordinate system. Experimental results indicated that only several test points were needed to establish the conversion relationship between the two coordinate systems, thereby obtaining the position of the person. The accuracy meets the demand of the stage performance.

Broadcasting voice is used to convey ideas and emotions. In the selection process of broadcasting and hosting professionals, the vocal timbre is an important index. The subjective evaluation method is widely used, but the selection results have certain subjectivity and uncertainty. In this paper, an objective evaluation method of broadcasting vocal timbre is proposed. Firstly, the broadcasting vocal timbre database is constructed on Chinese phonetic characteristics. Then, the timbre feature selection strategy is presented based on human vocal mechanism, and the broadcast timbre characteristics are divided into three categories, which include source parameters, vocal tract parameters, and human hearing parameters. Finally, the three models of hidden Markov model (HMM), Gaussian Mixture Model-General Background Model (GMM-UBM), and long short-term memory (LSTM) are exploited to evaluate the timbre of the broadcast by extracting timbre features and four timbre feature combinations. The experiments show that the selection of timbre features is scientific and effective. Moreover, the accuracy of the LSTM network using the deep learning algorithm in the objective evaluation of the broadcast timbre is better than the traditional HMM and GMM-UBM, and the proposed method can achieve about 95 accuracy rate in our database.