Search

In this letter, we developed novel Structure Enhanced Channel Attention Network (SECANet) for scene text image super‐resolution (STISR). The newly proposed SECANet integrates a group of Structure‐Enhanced Attention Modules to focus more on both local and global structural features in the character regions of text images. Moreover, we elaborately formulate a Dual‐Domain Contrastive Learning framework that integrates one pixel‐level contrastive loss and the other semantic‐level contrastive loss to jointly optimize the SECANet for generating more visually pleasing yet better recognizable high‐quality SR images without introducing any additional prior generators in both the training and testing stages, showing promising computational efficiency. Experimental results on the Textzoom dataset indicate that our method can achieve both decent performance in super‐resolving more impressive scene text images from low‐resolution ones and better recognition accuracy than other competitors. [ABSTRACT FROM AUTHOR]

The generation process of low‐light images is fundamentally complicated due to their multi‐factorial degradation. Many previous methods favorably developed a multi‐branch network for impressive performance. However, the complex structures of multi‐branch networks always incur large computational cost and hinders their applicability in those resource‐limited settings. To alleviate these issues, in this letter a novel lightweight multi‐branch network (LMBN) for low‐light image enhancement is proposed. Specifically, a multi‐branch module and a joint loss are framed to surmount the complex degradation factors of low‐light images. Furthermore, the multi‐branch network is implemented by several heterogeneous and shallow encoder‐decoder modules. The elaborately designed lightweight model facilitates lower computational complexity but maintains impressive enhancement quality. The experimental results performed on three popularly benchmark databases demonstrate that the proposed LMBN shows the superior performance over other state‐of‐the‐art methods, indicating a perfect balance between the model's performance and the computational efficiency. [ABSTRACT FROM AUTHOR]

In this Letter, the authors present a novel graph clustering-based method for crowd counting only using very limited labelled samples. Based on an intuitional observation that the distribution of low-level features of a specific scene containing the same or similar number of pedestrians are close to each other in the feature space, the authors adopt a first neighbour propagation (FNP) based clustering method to divide all unlabelled data into different groups. Next, an active sampling learning strategy that measures representativeness and diversity of the training data is used to obtain a few informative samples for annotation. Finally, the counts of those labelled informative samples are effectively propagated to predict the unlabelled samples in the constructed clusters by FNP-based clustering. The compelling results on two benchmark datasets demonstrate that the proposed method is not only effective to estimate crowd counts with very few labelled samples but also applicable to annotate a large number of unknown video frames for scene-specific crowd counting models. [ABSTRACT FROM AUTHOR]

Due to the issue of disorder, it is difficult to directly utilise a 2D convolutional neural networks to process 3D point clouds. Recently, PointNet can directly use 3D point sets as the input of convolutional neural networks and complete the processing of point clouds with multi‐layer perceptron (MLP) and symmetric functions. However, the use of MLP to compute the weight function ignores the problem of non‐uniformity sampling caused by the density of point set data. To address the above problem, based on the PointNet++ structure, a kernel density estimation based method is proposed to calculate the density level of the local point sets region under the optimal bandwidth selection principle, and the density re‐weighting of the weight function is developed to better fit the structure of local point clouds. In addition, the authors utilise the upsampling convolution operation to avoid duplicate storages and calculations, making the point cloud reconstruction more efficient. The experiments carried out the semantic segmentation on both the synthetic data and the real indoor scenes show that the proposed method is capable of obtaining promising semantic segmentation results. [ABSTRACT FROM AUTHOR]

Recently, a data‐driven approach from point cloud upsampling network (PU‐Net) has been used to upsample point set, making it from sparse to dense. However, PU‐Net first needs to downsample the point set before upsampling, and extract the point features of fixed‐level in the sampling area. The local features are not enough to extract, and then restoring its global representation will lead to point cloud coordinates to be inaccurate. In view of the insufficient local feature extraction, EdgeConv module is utilised to merge the points and the information in the local neighbourhood. In addition, a dense connection is proposed to maintain the effective transmission of information across the entire network. In this letter, EdgeConv and dense connection are combined into DenseConv that are nested for recursive use on PU‐Net to ensure the accuracy of the generation of point cloud. The experimental results show that the average error of the proposed method is reduced by about 5% on the PU‐Net dataset and about 10% on the ModelNet10 dataset, which verifies the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]