基于自适应模板更新与多特征融合的视频目标分割算法

Video object segmentation algorithm based on adaptive template updating and multi-feature fusion

针对SiamMask不能很好地适应目标外观变化,特征信息利用不足导致生成掩码较为粗糙等问题,本文提出一种基于自适应模板更新与多特征融合的视频目标分割算法。首先,算法利用每一帧的分割结果对模板进行自适应更新;其次,使用混合池化模块对主干网络第四阶段提取的特征进行增强,将增强后的特征与粗略掩码进行融合;最后,使用特征融合模块对粗略掩码进行逐阶段细化,该模块能够对拼接后的特征进行有效的加权组合。实验结果表明,与SiamMask相比,本文算法性能有明显提升。在DAVIS2016数据集上,本文算法的区域相似度和轮廓相似度分别为0.727和0.696,比基准算法提升了1.0%和1.8%,速度达到40.2 f/s;在DAVIS2017数据集上,本文算法的区域相似度和轮廓相似度分别为0.567和0.615,比基准算法提升了2.4%和3.0%,速度达到42.6 f/s。

In order to solve the problem that SiamMask cannot adapt to the change of target appearance and the lack of use of feature information leads to rough mask generation,this paper proposes a video object segmentation algorithm based on the adaptive template update and the multi-feature fusion. First of all, the algorithm adaptively updates the template using the segmentation results of each frame;secondly, the hybrid pooling module is used to enhance the features extracted in the fourth stage of the backbone network, and the enhanced features are fused with the rough mask;finally, the feature fusion module is used to refine the rough mask stage by stage, which can effectively combine the spliced features. Experimental results show that, compared with SiamMask, the performance of the proposed algorithm is significantly improved. On the DAVIS2016 data-set, the region similarity and contour similarity of this algorithm are 0.727 and 0.696, respectively, which is 1.0% and 1.8% higher than that of the benchmark algorithm, and the speed reaches 40.2 f/s. On the DAVIS2017 data-set, the region similarity and contour similarity of this algorithm are 0.567 and 0.615, respectively, which is 2.4% and 3.0% higher than that of the benchmark algorithm, and the speed reaches 42.6 f/s.