Efficient image blur detection via hierarchical edge guidance and region complementation

Abstract Blur detection is aimed to recognize the blurry pixels from a given image, which increasingly valued in vision-centered applications. Albeit great improvement achieved by recent deep learning-based methods, overweight model and rough boundary still pose challenges blur detection. In this paper, we propose Hierarchical Edge-guided Region-complemented Network (HER-Net) tackle above issues quest of favorable accuracy–complexity trade-off. First, novel olive-shaped pear-shaped inverted bottleneck structures based on large-kernel depth-wise convolutions build very concise architecture. Second, provoke exploit region-concerned edge-concerned morphological priors refine boundary. To end, reverse-region spatial attention mine complementary affinities between sharp regions so as enrich residual details around addition, an edge guide cues emphasize features related Both attentions are embedded into with hierarchical manners. Extensive experiments three benchmark datasets demonstrate that proposed method can achieve better performance using fewer parameters lower floating-point operations compared competitive methods. It proves efficiency effectiveness our task.