| There are two main implementation methods of the existing intelligent anti-blocking danmaku technology.One is to use the traditional pattern recognition method,which has poor segmentation effect,slow running speed,and poor comprehensive experience;the other is the semantic recognition method based on deep learning.Segment the network model and segment the image pixel by pixel.Compared with the traditional method,this method has greatly improved the accuracy and speed.However,due to the large number of complex scenes in video content,the performance of semantic segmentation networks is limited,and the problem of poor segmentation quality still exists in practical applications.Aiming at the problems existing in existing methods such as blurred segmentation edges,missing or incorrect segmentation of small objects and occluded objects,and slow running speed,this dissertation conducts the following research:(1)Aiming at the problems of poor edge processing,flickering danmaku,and poor segmentation of small targets and occluded targets,this dissertation designs an end-to-end intelligent anti-blocking danmaku method based on queries.On the one hand,the query vector is enhanced by introducing a multi-head self-attention mechanism,and the query vector can flow between different tasks through the cascaded network structure;on the other hand,a target tracking module is designed for video with frame correlation,through Compare the similarity of instances across frames to fuse effective information between frames.In order to verify the network performance,subjective and objective evaluations were carried out on two data sets.Compared with other methods,the method in this dissertation has certain improvements in various indicators,and a better experience in subjective perception.(2)In order to further improve the performance of the intelligent anti-blocking danmaku method,Swin-Transformer is used as the backbone network of the network,but the Transformer model lacks a spatial inductive bias.Based on this,the Adapter module is introduced to enhance Swin-Transformer.The Adapter module can supplement the space inductive bias without changing the original Swin-Transformer structure,and further enhance the Transformer’s ability to process images.The rationality of introducing the Adapter module is verified by conducting horizontal and vertical comparative experiments on the COCO data set and the YouTube-VIS data set,and evaluating from the two dimensions of objective and subjective.(3)In view of the slow processing speed of the existing intelligent anti-blocking danmaku method,the complexity of the self-attention mechanism(Self-Attention)in the Transformer module is optimized to reduce its complexity to linear complexity.Through experiments on two data sets,the parameter quantity,accuracy and processing speed are compared and analyzed.The results show that the proposed optimization scheme can achieve the effect of reducing the parameter quantity and improving the model processing speed while ensuring the accuracy. |
