Researches On Defocus Blur Detection Algorithms Based On Deep Learning

Defocus blur detection is an important component in computer vision,with the goal of detecting defocus blurred and clear regions of a given image and achieving effective segmentation of focused clear and defocus blurred regions of an image.Defocus blur detection algorithms can assess the performance of image denoising and deblurring algorithms;integrating such algorithms into image denoising and deblurring algorithms can improve image visual quality while reducing computational time.Other applications that use defocus blur detection include automatic focus,image super-resolution reconstruction,salient object detection,and so on.In recent years,convolutional neural networks have been widely used in various computer vision tasks because of their powerful feature extraction capabilities.Similarly,the application of deep learning-based methods in the field of defocus blur detection is becoming more and more widespread.Although the deep learning-based algorithms have achieved higher performance and significant improvements compared with traditional methods,there are still problems such as defocus blur detection in lowcontrast regions,cluttered boundaries of focused objects,defocus blur detection in homogeneous regions,and complementary fusion between cross-level features.Therefore,this paper addresses the above problems by conducting an in-depth study of deep learning-based defocus blur detection algorithms and designing two different network models.The main innovative research work and results of this paper are summarized as follows:(1)Hierarchical edge-aware network for defocus blur detectionTo address the problems of low-contrast focal region detection and coarse object boundaries in defocus blur regions.In this paper,we propose a hierarchical edge-aware network to address the aforementioned issues;to the best of our knowledge,this is the first trial to develop an end-to-end network with edge awareness for defocus blur detection.To capture boundary information,we design an edge feature extraction network,and a hierarchical interior perception network is used to generate local and global context information,which is useful for detecting low contrast focal regions.Furthermore,a hierarchical edge-aware fusion network is proposed to fuse edge information and semantic features hierarchically.The fused features can generate more accurate boundaries by leveraging the rich edge information.The method is compared with 12 defocus blur detection algorithms on two publicly available datasets,and the results of the experiment indicate that the proposed model performs better than state-of-the-art methods in both qualitative and quantitative evaluation.(2)Defocus blur detection via adaptive cross-level feature fusion and refinementTo address the problems of unsatisfactory defocus blur detection in homogeneous regions and the complementary information between cross-level features cannot be fully utilized.We propose a novel model via adaptive cross-level feature fusion and refinement for DBD,which mainly consists of adaptive cross-level feature fusion module,cross-level feature refinement module,and training dataset augmentation.Specifically,adaptive cross-level feature fusion module is used to adaptively discriminate different levels of features and effectively aggregate them,which leverages an adaptive fusion mechanism with self-learning weights.Moreover,we design the cross-level feature fusion module to refine the cross-level feature information at the decoder stage from coarse to fine.Furthermore,we extract the homogeneous region patches of the training images by utilizing the Laplace filter,which aims to improve the model robustness for homogeneous regions.It is found that our two proposed algorithms significantly outperform the other 13 compared algorithms under different evaluation metrics for the two widely used datasets.