基于超像素图割的多类别弱标注强化算法

Multi-classes weak labeling enhancement algorithm based on hyper-pixel graph cut

为解决基于深度学习的图像语义分割逐像素制作语义标签训练集耗时耗力的问题,提出一种便捷的基于超像素图割的多类别弱标注强化算法。在弱标注框内自适应提取超像素,采用交互式涂鸦结合超像素扩充前景背景采样点;根据框内采样点对高斯混合模型参数进行初始化;迭代更新参数,使用最小割算法对像素点进行分类,实现像素级强标注。实验结果表明,在保证标注精度的前提下,该方法较传统人工与Grabcut算法在标注上具有较大效率优势,对服装图像重新标注并作为全卷积网络训练集,达到与原始数据集相近的分割精度。

To solve the time-consuming and labor-intensive problems in making large-scale semantic tag training set based on deep learning image semantic segmentation,a convenient multi-classes weak tagging enhancement algorithm based on superpixel graph cuts was proposed. Superpixels in the weak rectangle were adaptively generated and the interactive scribbles were used combined with superpixels to expand the foreground background sampling points. The Gaussian mixture model parameters were initialized according to the foreground and background sampling points in the box. The Gaussian mixed model parameters were iteratively updated,and the minimum cut algorithm was used to classify pixels to obtain multi-category pixel-level strong annotation of images. Experimental results show that,the efficiency of the annotation is guaranteed,and compared with the traditional manual and Grabcut algorithm labeling,the proposed method has greater efficiency advantage. The clothing image is re-labeled and used as a full convolutional network training set to achieve similar segmentation accuracy in terms of the original data set.