基于DCTMask改进的实例分割方法分析

Analysis of Improving Instance Segmentation Network Based on DCTMask

阐述在实例分割任务中,Mask RCNN是非常热门的主线网络,但Mask R-CNN使用高分辨网格预测掩码则会大大增加训练的复杂性,低分辨率网格预测掩膜捕获细节不足。引入离散余弦变换方法,将高分辨率二值网格掩码编码成紧凑的向量,从而降低复杂度,提升效率。在神经网络常见的几种激活函数中,尤其是ReLU激活函数也存在高深度的网络中信息保留不足的问题,因此引入Swish激活函数,改进更深层网络中的激活函数的表现。

This paper expounds that Mask R-CNN is a popular network in the task of instance segmentation,but in the past,using high-resolution grid prediction masks has greatly increased the complexity of training Mask R-CNN,and low-resolution grid prediction masks have inadequate detail capture.The introduction of DCT methods allows for compact encoding of high-resolution binary binary grid masks as vectors,which reduces complexity and improves efficiency.When dealing with deep networks in which several commonly used activation functions are frequently applied,it often presents insufficient information preservation problems.Therefore,this chapter introduces Swish activation functions to improve the performance of deeper networks.