Prior-Induced Information Alignment for Image Matting

Image matting is an ill-posed problem that aims to estimate the opacity of foreground pixels in image. However, most existing deep learning-based methods still suffer from coarse-grained details. In general, these algorithms are incapable felicitously distinguishing degree exploration between deterministic domains ( e.g. certain FG and BG pixels) undetermined uncertain in-between pixels), or inevitably lose information continuous sampling process, leading a sub-optimal result. this paper, we propose novel network named Prior-Induced Information Alignment Matting Network (PIIAMatting), which can efficiently model distinction pixel-wise response maps correlation layer-wise feature maps. It mainly consists Dynamic Gaussian Modulation mechanism (DGM) strategy (IA). Specifically, DGM dynamically acquire domain map learned prior distribution. The present relationship variation convergence process during training. On other hand, IA comprises Match Module (IMM) Aggregation (IAM), jointly scheduled match aggregate adjacent features adaptively. Besides, also develop Multi-Scale Refinement (MSR) module integrate multi-scale receptive field at refinement stage recover fluctuating appearance Extensive quantitative qualitative evaluations demonstrate proposed PIIAMatting performs favourably against state-of-the-art image on Alphamatting.com , Composition-1 K Distinctions-646 dataset.