基于不同场景下三分图扩展的图像抠图算法

Image matting based on the trimap expanded in different scenarios

图像抠图是计算机视觉的关键任务之一。在图像抠图中,粗糙三分图的扩展可以提高抠图质量、节省成本,因此被广泛应用于图像抠图。但基于三分图扩展的抠图方法不能同时应用在前景不透明和前景半透明图像场景。因此,在使用语义特征扩展背景的基础上,引入了像素间的图像特征、类别特征感知不同透明度的输入图像,对不同场景的三分图前景分别扩展,将扩展后的三分图用于图像抠图。此外,在Composition-1k数据集上与目前的三分图扩展方法效果进行了对比分析。结果表明,在综合评价指标上,提高了不同场景下的低质量三分图的抠图质量。

Image matting is one of the key tasks in computer vision.The expansion algorithm of the trimap can improve the quality of matting and save costs.This trimap expansion method is widely used.However,the matting method based on the expansion of the trimap cannot simultaneously handle foreground opaque and foreground translucent image scenes.Therefore,on the basis of using semantic features to expand the background,pixel-level image features and category features are introduced to perceive input images with different transparency levels.The foreground of the trimap is expanded separately for different scenes,and the expanded trimap is used for image matting.In addition,a comparative analysis of the effect of the current trimap expansion method is conducted on the Composition-1k dataset.The results show that the quality of matting of low-quality trimap under different scenes is improved in terms of comprehensive evaluation indicators.