HPMatte: Hybrid Semantic Pyramid Architecture for Image Matting

Image matting is a critical yet complex task in computer vision. Recent advancements in deep learning have significantly improved performance in this domain. However, the majority of these approaches rely on a trimap as an auxiliary input, which restricts their usability in practical applications. While some methods have attempted to remove the dependency on trimaps, their matting quality generally lags behind that of trimap-assisted models. The absence of trimap guidance often leads to foreground-background ambiguities and imprecise details in transition areas. To address these limitations, we introduce HPMatte, a novel matting framework combining Transformer and CNN (Convolutional Neural Network) architectures. This hybrid model achieves high-quality matting results without the requirement of a trimap input. The key component of our model is the Pyramid Semantic Block (PSB), which extracts features at different scales, fuses information from different resolutions, and preserves fine details of the foreground. This enables high-precision matting of natural images. Additionally, A background dataset called GBG-10k is introduced which enhances the diversity of existing matting datasets. The approach is evaluated by using two well-known benchmark datasets, i.e., AM-2k and P3M-10k. The experimental outcomes highlight the advantages of HPMatte compared to existing approaches.