计算机工程

• •    

基于CornerNet-Saccade的手部分割算法研究

  

  • 发布日期:2021-01-05

Research on Hand Segmentation Algorithm Based on CornerNet-Saccade

  • Published:2021-01-05

摘要: 手部分割技术在人机交互与医疗健康领域有巨大的应用前景,由于手部形态多变、分割背景复杂等因素,使得分割效率很难提高。为提 高分割效率,本文在 CornerNet-Saccade 模型的基础上进行改进,构造了一种基于扫视机制的手部分割模型。该模型模拟人眼观察物体时先扫视再 仔细观察的方法,降低了要处理的像素数量。初步判断手部位置后,在不同尺度特征图中添加掩码分支,完成精细分割任务。另外,为进一步改 善模型复杂度高的问题,引入线性瓶颈结构完成了轻量化操作。实验表明,该模型在 Egohands 数据集上平均交并比(mIOU:Mean Intersection over Union )达到了 88.4%,优于 RefinNet、U-Net 等主流方法。轻量化处理后,与原模型相比,平均交并比降低了 2.2%,但参数量只有原先的 44.9%。

Abstract: Hand segmentation technology has huge application prospects in the fields of human-computer interaction and medical and health. Due to factors such as variable hand shapes and complex segmentation background, it is difficult to improve the segmentation efficiency. In order to improve the efficiency of segmentation, this paper improves on the CornerNet-Saccade model and constructs a hand segmentation model based on saccade mechanism. This model simulates the method of scanning and then carefully observing the human eye when observing an object, reducing the number of pixels to be processed. After preliminary judgment of the hand position, mask branches are added to the feature maps of different scales to complete the fine segmentation task. In addition, in order to further improve the problem of high model complexity, a linear bottleneck structure is introduced to complete the lightweight operation. Experiments show that the model achieves 88.4% mIOU on the Egohands dataset, which is better than mainstream methods such as RefinNet and U-Net. After lightweight processing, compared with the original model, the average intersection ratio is reduced by 2.2%, but the parameter amount is only 44.9% of the original model.