Video object segmentation method based on self-paced weak supervised learning

Abstract

The invention provides a video object segmentation method based on self-paced weak supervised learning. A self-paced learning algorithm is embedded into a depth neural network, under the guidance of the thought of weak supervised learning, a whole system learns target concepts from the easier to the more advanced, the network obtained by learning with the training process becomes complex, the ability of the network to deal with problems is gradually increased, and finally, an accurate video object segmentation result is obtained. The invention utilizes the advantages of the self-paced learning algorithm and the deep neural network model comprehensively, has higher segmentation accuracy, and shows better robustness when processing video data of different scenes.

Description

technical field [0001] The invention belongs to the field of computer vision algorithm research, and specifically relates to a method for completing video object segmentation tasks by combining a self-paced learning method into a deep neural network under the category of weakly supervised learning. Background technique [0002] In recent years, the rapid development of social media and video sharing websites has made the demand for video processing more and more intense, and the use of weakly supervised learning algorithms for video object segmentation has great application value. [0003] There have been a lot of work on video object segmentation methods, such as Key-Segments for Video Object Segmentation proposed by Yong Jae Lee et al. in 2011 and Video Object Segmentation through Spatially Accurate and Temporally DenseExtraction of Primary proposed by Dong Zhang et al. in 2013 Object Regions, these existing methods generally follow the following working framework: for a s...

AI Technical Summary

Problems solved by technology

[0004] Although the above framework has achieved good results, it still has some limitations: First, in terms of overall design, this type of method is only composed of many serial processing units, rather than end-to-end planning of the problem, this design The method relies too much on the researcher's professional knowledge and own experience, which may limit the performance of the system

Secondly, most existing methods process each frame of video separately during the learning process, without considering the information provided by other video frames under the same semantic category

Finally, the above framework needs to use negative sample data during the training process. The uncertainty in the quantity and quality of negative sample data may lead to instability in the final performance of the method.

Images