Defocus blur detection via adaptive cross-level feature fusion and refinement

Convolutional neural networks have achieved competitive performance in defocus blur detection (DBD). However, due to the different receptive fields of different convolutional layers, there are distinct differences in the features generated by these layers, and the complementary information between cross-level features cannot be fully utilized. Besides, there are still challenges to be solved in homogeneous regions. To tackle the above issues, we focus on both homogeneous region dataset augmentation and model design to propose a novel DBD model based on adaptive cross-level feature fusion and refinement. Specifically, in terms of homogeneous region dataset enhancement, a Laplace filter is used to extract the homogeneous region image patch of the training image to realize homogeneous region image augmentation, which improves the robustness of the model for DBD in the homogeneous region; in terms of model design, we propose an adaptive fusion mechanism with self-learning weights and design the adaptive cross-level feature fusion module, which adaptively discriminates between different levels of features and fuses them step-by-step. In addition, we design the cross-level feature refinement module and embed it into the network, which captures the complementary information of the cross-level features, and refines cross-level feature information from coarse to fine in the decoder stage. Experimental results on two commonly used datasets show that the proposed method outperforms 13 state-of-the-art approaches.