引用本文
  •    [点击复制]
  •    [点击复制]
【打印本页】 【在线阅读全文】【下载PDF全文】 查看/发表评论下载PDF阅读器关闭

←前一篇|后一篇→

过刊浏览    高级检索

本文已被:浏览 280次   下载 281 本文二维码信息
码上扫一扫!
基于注意力机制的实例分割算法
张声传,喻松林,纪荣嵘
0
(厦门大学信息学院,厦门 361005)
摘要:
实例分割作为计算机视觉领域极具挑战性的任务之一,要求在图像分类的基础上为每一个物体生成像素级别的分割掩码。业界主流方案可分为自上而下和自下而上两种范式,自上而下范式又可分为双阶段分割和单阶段分割。单阶段分割方案为了提高推断速度,往往使用全图卷积操作取代双阶段分割方案中先检测后分割的策略。然而,卷积网络的平移不变性使得同一种类的不同实例提取到的特征相似,仅靠全图卷积难以进行区分,从而导致单阶段分割方案精度下降。针对单阶段分割精度降低的问题,提出了一种注意力机制,该机制在特征图每个位置的特征向量上进行点积运算,并将运算结果作为新的特征图,同一位置点积结果最大化,不同位置点积结果最小化,以丰富特征图中不同实例的差异信息。通过注意力机制使得单阶段分割方案中的全图卷积操作能更好地区分同一种类的不同实例,从而生成高质量分割掩码。在公开数据集上进行实验,验证了所提方法的有效性。
关键词:  实例分割  单阶段分割  注意力机制
DOI:
基金项目:国家杰出青年科学基金(62025603);国家自然科学基金(U1705262,62072386,62072387,62072389, 62002305,61772443,61802324,61702136);广东省基础与应用基础研究基金(2019B1515120049);中央高校基本科研业务费(20720200077,20720200090,20720200091)
Research on Instance Segmentation via Attention Mechanism
ZHANG Sheng-chuan,YU Song-lin,JI Rong-rong
(School of Information, Xiamen University, Xiamen 361005, China)
Abstract:
Instance segmentation, as one of the most challenging tasks in the field of computer vision, requires the generation of a pixel-level segmentation mask for each object based on image classification. The mainstream solutions in the industry can be divided into top-down and bottom-up paradigms. The top-down paradigm can be divided into two-stage segmentation and one-stage segmentation. In order to improve the speed of inference in one-stage segmentation schemes, full-image convolution operations are often used instead of the strategy of first detection and then segmentation in the two-stage segmentation scheme. However, the translation invariance of the convolutional network makes the features extracted by different instances of the same type similar, and it is difficult to distinguish them only by the whole image convolution, which leads to the reduction of the accuracy of the one-stage segmentation scheme. Aiming at the problem of one-stage segmentation accuracy reduction, an attention mechanism is proposed. This mechanism performs dot product operation on the feature vectors of each position of the feature map, and uses the result of the operation as a new feature map. The result of the dot product at the same position is maximized and that at the different location is minimized, so that to enrich the difference information of different instances in the feature map. Through the attention mechanism, the full-image convolution operation in the one-stage segmentation scheme can better distinguish different instances of the same type, and generate high-quality segmentation masks. Tests on public data sets verify the effectiveness of the proposed method.
Key words:  Instance segmentation  One-stage segmentation  Attention mechanism

用微信扫一扫

用微信扫一扫