FANet: Feature attention network for semantic segmentation

Semantic segmentation based on scene parsing specifies a category label for each pixel in the image. Existing neural network models are useful tools for understanding the objects in the scene. However, they ignore the heterogeneity of information carried by individual features, leading to pixel classification confusion and unclear boundaries. Therefore, this paper proposes a novel Feature Attention Network (FANet). Firstly, the adjustment algorithm is presented to capture attention feature matrices that can effectively cherry-pick feature dependencies. Secondly, the hybrid extraction module (HEM) is constructed to aggregate long-term dependencies based on proposed adjustment algorithm. Finally, the proposed adaptive hierarchical fusion module (AHFM) is employed to aggregated multi-scale features by learning spatially filtering conflictive information, which improves the scale invariance of features. Experimental results on popular Benchmarks (such as PASCAL VOC 2012, Cityscapes and ADE20K) indicate that our algorithm achieves better performance than other algorithms.