基于YOLACT与Transformer相结合的实例分割算法研究

Research on instance segmentation algorithm based on YOLACT and Transformer

摘要:

为提高单阶段实例分割的检测精度和改善小目标漏检、错检情况，提出一种基于YOLACT改进的YOLACTR算法。该算法首先利用CNN与Transformer相结合，设计一种新的头部预测网络，对特征进一步提取，并使用双向注意力来关联同一实例的掩码信息并区分不同实例之间的掩码特征，注重特征点周围的关联信息，使得检测框的预测更加准确；然后利用多级上采样和设计的CS注意力模块结合形成掩码分支，使其融入多种不同尺度信息，并利用CS注意力来关注不同的尺度信息。在MS COCO数据上，YOLACTR算法与YOLACT算法相比，其边框和掩码检测精度分别提升了7.4%和2.9%，在小目标检测上分别提升了18.9%和13.5%。实验表明，YOLACTR算法可以在多目标复杂场景下，提升检测和分割精度以及分类的准确度，改善了小目标和重叠目标漏检、错检的问题。

Abstract:

In order to improve the segmentation accuracy of single stage instance segmentation and improve the small target detection, an improved YOLACTR algorithm is proposed based on YOLACT algorithm. The algorithm first uses the combination of CNN and Transformer to design a new head prediction network to further extract features, and uses two-way attention to correlate the mask information of the same instance and distinguish the mask features between different instances. It pays attention to the correlation information around the feature points, making the prediction of the detection box more accurate. Then the mask branch is formed by the combination of multi-level up sampling module and the designed CS attention module, which integrates a variety of different scale information, and uses CS attention module to pay attention to different scale information. On the MS COCO, compared with YOLACT algorithm, YOLACTR algorithm improves the detection accuracy of Box and Mask by 7.4% and 2.9% respectively, and improves the detection accuracy of small targets by 18.9% and 13.5% respectively. Experiments show that YOLACTR algorithm can improve the accuracy of detection, segmentation and classification in multi-target complex scenes, and improve the problems of missed detection and false detection of small targets and overlapping targets.