Defocus Blur detection via transformer encoder and edge guidance

Defocus blur detection (DBD) aims to separate blurred and unblurred regions for a given image. Benefiting from the powerful extraction capabilities of convolutional neural networks (CNNs), deep learning based defocus blur detection has achieved a remarkable progress compared with traditional methods. However, due to the limited local receptive field of CNNs, it is difficult to achieve satisfactory results in the detection of the low-contrast focal regions. Besides, the output maps of the most of previous works have coarse object boundaries and background clutter. In this paper, we propose a hybrid CNN-Transformer architecture with an edge guidance aggregation module (EGAM) and a feature fusion module (FFM) for DBD. In our knowledge, this is the first study to utilize a transformer encoder for DBD to capture the global context information. Additionally, an edge extraction network (EENet) is adopted to obtain local edge information of in-focus objects. To effectively aggregate local edge information and global semantic features, three EGAMs are integrated into an edge guidance fusion network (EGFNet). Benefiting from the rich edge information, the fused features can generate more accurate boundaries. Finally, three FFMs are cascaded as a hierarchical feature aggregation network (HFANet) to hierarchically decode and refine the feature maps. Experimental results on three widely used DBD datasets demonstrate that the proposed model outperforms the state-of-the-art approaches.