基于胶囊的小样本语义分割方法

Method of few-shot semantic segmentation based on capsules

语义分割是指对图像进行像素级的分类并且对图像中的每个像素都给定一个类标记的图像分割技术,是计算机视觉领域的主要研究方向之一。随着深度学习的发展,基于深度神经网络的语义分割方法发展迅速,并实现了超越传统语义分割方法的精度。然而,基于深度学习的语义分割的实现,需要大量的有标记的训练数据支持。在许多应用场景中,得到大量的标记数据是困难的。因此,基于小样本的语义分割方法被提出,以解决这类问题。当前的小样本语义分割方法主要是通过预训练的分类网络提取高维特征,然后使用高维特征的余弦相似度图为引导,对查询图像的目标区域进行分割。但是这些方法都面临着相似度图不清晰导致分割精度不高的问题。为此,文中设计了一个胶囊模块,用于生成更清晰的余弦相似度图。此外,为更好地利用多个支持图像中包含的语义信息,除了使用交叉熵损失函数外,还使用了一个新的边缘损失函数。在国际计算机视觉竞赛数据集(PASCAL-5^(i))上的试验结果显示:小样本(5个样本)语义分割方法精度平均交并比(mean intersection over union,mIoU)达到61.0%。

Semantic segmentation refers to image segmentation technology that classifies images at the pixel level and assigns a class tag to each pixel in the image.It is one of the main research directions in the field of computer vision.With the development of deep learning,CNN-based semantic segmentation methods have developed rapidly and have achieved better performance than traditional semantic segmentation methods.However,the CNN-based semantic segmentation requires a large amount of labeled training data.In many application fields,it is difficult to obtain a large amount of labeled data.Therefore,the few-shot semantic segmentation method is proposed to solve such problems.The current few-shot semantic segmentation method mainly uses a pre-trained classification network to extract high-level features and then uses the cosine similarity map of the high-level features as a guide to segment the target region of the query image.However,these methods are faced with the problem of unclear similarity maps leading to faulty segmentation results.To tackle this problem,we proposed a capsule module to generate a clearer cosine similarity map.To make better use of the semantic information contained in multiple supporting images,we used a new margin loss function along with the cross-entropy loss function to optimize our model.The experimental results on the public data set PASCAL-5^(i) show that our 5-shot semantic segmentation method has achieved 61.0%mIoU.