Defocus Blur Detection via Multi-stream Bottom-Top-Bottom Fully Convolutional Network

Convolutional Neural Network for Extracting 3D Point Clouds of Fibrous Web From Multi-Focus Images

“…The smaller MAE, the higher the accuracy. To compare the performance of different algorithms, we used three stateof-the-art in-focus region detection models: BTBNet [43], LBP [47] and MGF [46]. Because most of the objects are fiber edges in the dataset, the state-of-the-art edge detection models, such as BDCN [42], RCF [41], HED [40] and U-Net [24], were also selected to compare with the MIDN.…”

“…However, the experiments showed that this method cannot distinguish the low-contrast focus regions. Zhao et al [43] designed a multi-stream network (BTBNet) to detect in-focus regions. The BTBNet combines multiple convolutional layers to compose streams, and utilizes the streams to extract features in different scales.…”

Self-supervised blur detection from synthetically blurred scenes

“…By extensively evaluating the learned model on the two largest publicly available pixel-wise annotated real blur localization datasets [11,35], we show that our approach generalizes adequately for both blur types, improving the performance of all the considered classic approaches and that of recent fully supervised deep CNN-based ones that require large human-labeled training sets. Experimental results show that the proposed solution can be successfully employed to train a deep blur segmentation model, either without the need for any specific blur localization dataset or by making use of a very reduced set of images exhibiting real blur degradations.…”

“…More recently, supervised learning-based approaches [25], and particularly those based on the use of Convolutional Neural Networks (CNN), have shown enormous potential for tackling tasks that require a dense, per-pixel prediction, such as semantic segmentation [27,28], instance segmentation [29] or crowd counting via density map estimation [30]. Blur segmentation can also be viewed as one of such dense prediction tasks, and several works have already explored this approach, either for predicting both types [31,1,32] or defocus only blur [33,34,35,36]. Nevertheless, the performance gain obtained by these fully-supervised, CNN-based approaches trained end-toend is relatively modest when compared to gains in other fields.…”

“…The main bottleneck which hinders the full power of CNNs in their application to the field of blur localization is the absence of large enough datasets with pixel-wise annotations. Several recent efforts have partially addressed this void [11,1,35,36], but these manual annotations are scarce, labour intensive and costly to acquire.…”