基于Generalized Region Loss的代价函数及在图像分割中的应用

A Cost Function Based on Generalized Region Loss and its Application in Image Segmentation

摘要:

针对图像分割中的困难样本，提出了一种对像素区域细分计算的Generalized Region Loss的新的代价函数；首先通过引入一项参数，改变了以往代价函数主要通过设置权重或Focal等关注困难样本的方法，其次通过对标签图像和预测图像进行区域划分，并且对划分四区域的困难样本分类关注，最后分别计算其四区域绝对损失，进而进行加权组合；为验证算法性能，使用CamVid数据集作为实验数据，该代价函数在FCN和U-Net两种图像分割网络上得到验证，同当前图像分割领域常用的12种代价函相比，IoU指标分别提高1.93%和2.99%，由此证明此代价函数优于大多数图像分割代价函数；最终实验结果表明，提出的基于像素区域细分计算的代价函数能够有效提高图像分割精度，为图像分割的研究提供借鉴。

Abstract:

Aiming at the difficult samples in image segmentation, a new cost function of generalized region loss is proposed; Firstly, a parameter is introduced to change the previous method of focusing on difficult samples by setting weights or focal in the cost function. Secondly, the label image and the prediction image are divided into regions, and the difficult samples divided into four regions are classified and focused. Finally, the absolute losses of the four regions are calculated respectively, and then the weighted combination is carried out; In order to verify the performance of the algorithm, the CamVid data set is used as experimental data. The cost function is verified on the FCN and u-net image segmentation networks. Compared with the 12 cost functions commonly used in the current image segmentation field, the IOU index is increased by 1.93% and 2.99% respectively, which proves that this cost function is superior to most image segmentation cost functions; The final experimental results show that the proposed cost function based on pixel segmentation can effectively improve the accuracy of image segmentation and provide a reference for the research of image segmentation.