【Portrait分割】BANet:Boundary-Aware Network for Fast and High-Accuracy Portrait Segmentation

文章目录

Abstract
1、Background
2、Method of BANet
- 2.1、Network Architecture
- - Boundary Feature Mining Branch
- 2.2、Loss Function

Abstract

与通用语义分割任务相比，Portrait分割需要更高的精度和更快的速度。

Boundary-Aware Network (BANet) 选择性提取边界区域的细节信息获得高质量的分割结果，且可做到实时分割（ ≥ 25FPS）

同时BANet设计了一个refine loss对网络中的图像级梯度信息进行监督。

BANet is an efficient network with only 0.62MB parameters, it achieves 43 fps on 512 × 512 images with high-quality results which are finer than annotations.

1、Background

精细的边界细节信息丢失问题主要有以下两个原因造成：

一方面，深度学习模型的性能非常依赖于训练数据。然后target通常用polygons来标注或通过KNN-matting生成。故像头发丝这种极其精细的边界细节很难被标注出来。

另一方面，传统语义分割任务主要解决复杂场景中的intra-class consistency and the inter-class distinction问题。Portrait分割属于二分类问题，传统语义分割模型并不合适。

2、Method of BANet

In the task of portrait segmentation, no-boundary area needs a large receptive field to make prediction with global context information, while boundary area needs small receptive field to focus on local feature contrast. Hence these two areas need to be treated independently.

In this paper, we propose a boundary attention mechanism and a weighted lossfunction to deal with boundary area and no-boundary area separately.

2.1、Network Architecture

Semantic Branch: 获得比较大的感受野，提高对非边界区域的分割。channel数最大仅为64;

Fusion Part: 采用BiSeNet中的FFM模块，channel attention mechanism

Boundary Feature Mining Branch

semantic branch的输出先通过1 × 1 conv映射到一个通道，然后再上采样到原图大小，作为boundary attention map。

BA loss引导boundary attention map定位边界区域。

Extraction of semantic boundary forces the network to learn a feature with strong inter-class distinction ability.

BA loss的target并不需要手动标注。先使用Canny边缘检测器检测portrait annotation（ground truth），然后检测结果作为BA loss的target。实现可参考：Canny

最后，输入图像与attention map拼接得到一个4维的特征图。

2.2、Loss Function

未完待续…