An end-to-end face parsing model using channel and spatial attentions. (15th March 2022)

• Record Type: Journal Article
• Authors: Kim, Hyungjoon; Kim, Hyeonwoo; Cho, Seongkuk; Hwang, Eenjun
• Is Part Of: Measurement. Volume 191(2022)
• Publication Date: 2022-03-15
• Subjects: Face parsing -- Attention mechanism -- Image segmentation; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.110807
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: Face parsing is the labeling of every pixel in a face image, and an image segmentation technique can be used. Attention mechanisms have improved the performance of convolutional neural networks, and the image segmentation field has also advanced by introducing these techniques. The new attention block, consisting of a channel attention block and a spatial attention block, accurately extracts the shape of facial components from an input image and assigns labels to the region of components. The channel and spatial attention blocks are designed to emphasize their strengths and complement their weaknesses. Abstract: Facial image parsing requires accurate extraction of facial components and features, and image segmentation can be used. Recently, various attention mechanisms showed excellent performance in segmentation by extracting features based on spatial and channel relationships for input images. In this paper, we propose a new face parsing technique using an attention block that combines the spatial attention block and the channel attention block to effectively utilize their functions. In this process, we improve the structure of the two blocks to compensate for their weaknesses. The attention block extracts features related to the shape of facial components from spatial relationships and concentrates on more important channels from correlation among channels. We built several segmentation models using the proposed block and compared their performance with well-known segmentation models. Experimental results showed that our combined block-based model can improve the segmentation accuracy by more than 5% in F1 score compared to other models.