Although video deraining technology has achieved great success in recent years, extracting spatiotemporal feature representations across the domains of spatial and temporal in successive frames, then performing spatial and temporal modeling, and restoring high-quality deraining videos with rich details are still challenging tasks. In this paper, we use the hybrid Transformer for the first attempt in video rain removal tasks, and propose a novel video deraining network based on hybrid transformer (VDN-HT) to aggregate global and local representations to accomplish video deraining. In the feature extraction process, we propose to use a U-shaped structure based on serial Transformer blocks to extract shallow local features, deep global features and global dependencies, and then adaptively aggregate them to obtain rainy video features with rain streaks of different directions and densities. In order to better model spatiotemporal relationships, the VDN-HT uses the Transformer’s long-range and relational modeling abilities to obtain the features of spatial and the correlations of temporal between continuous video frames to achieve multi-frame alignment. For ensuring the global-local consistency of the reconstructed frames, we design a global-local reconstruction module composed of Transformer and convolutional neural network (CNN) in parallel to aggregate global and local information to better reconstruct each frame. In addition, the proposed gating-based refinement module and color loss effectively retain the details and color information after removing rain streaks. Extensive experiments on NTURain, RainSynLight25 and RainSynHeavy25 datasets have shown that the VDN-HT can handle many types of rainy videos and perform better than previous methods.

Tiny objects often have a small proportion of pixels in the image, leading to significant differences in the number of positive and negative samples and the lack of feature information. Accurately determining the position and category of tiny objects remains a huge challenge for object detection research. Therefore, we design an Adaptive Sample Assignment Strategy(ASAS) and tiny object focusing enhancement module to solve the above two problems. Specifically, starting from the study of positive and negative sample selection and balance strategies for tiny objects, we construct a lightweight Object Existence Probability Determination Network (OEPD/Net) to focus on the areas where tiny objects exist, and achieve adaptive assignment and balance of samples. A top/down, layer by layer focusing enhancement module is designed to effectively enhance the propagation ability of high/level semantic information for tiny objects. The above two solutions have excellent generalization and migration capabilities and can be applied to any stage and two-stage object detection network, effectively enhancing TOD performance. Finally, this article provides a performance analysis of detection performance the detection network based on the OEPD/Net output results, and demonstrates the effectiveness of the proposed OEPD-Net and focusing enhancement module through extensive experiments on a public dataset.

Efficient and accurate point cloud feature extraction is crucial for critical tasks such as 3D recognition and semantic segmentation. However, existing global feature extraction methods for 3D data often require designing different models for different input types (point clouds, voxels, and maps). This article proposes an efficient plug-and-play non-iterative clustering method (NICM) to establish a unified point cloud global feature extraction paradigm suitable for any input type to solve the above problems. The core idea of the NICM is to construct the connection between a single point and other global points based only on the cosine similarity between center points to achieve global feature extraction, which has linear complexity characteristics and can be combined with any existing feature extraction model. Additionally, to better integrate the features extracted by NICM and the original model, this article designs an adaptive feature fusion module is designed based on the gate unit, which retains similar features and effectively fuses dissimilar features based on their importance to downstream tasks. We have applied our method to downstream tasks such as point cloud recognition, part segmentation and scene segmentation. Sufficient experiments have proven that our method can provide comprehensive and robust features for the original model, and effectively improve the performance of downstream tasks.

Abstract Automatic and accurate instance segmentation of teeth can provide important support for computer‐aided orthodontic work. Traditional methods for tooth segmentation studies often ignore the rich structural features of teeth. Capturing the complete and accurate geometry as well as morphological details of a single tooth remains a challenge for current tooth segmentation studies. In this article, a new tooth segmentation deeplearning network based on capturing dependencies and receptive field adjustment in cone beam computed tomography (CBCT) is proposed to achieve automatic and accurate instance segmentation of dental CBCT data. The method acquires coarse‐level features of tooth and accurate tooth centroids in the first stage, and acquires the instance information and spatial position localization of the tooth. The encoding process in the second stage of the network introduces a guidance module for obtaining tooth geometry information based on a 3D self‐attention mechanism to capture dependencies in CBCT. The proposed tooth feature integration module is based on multiscale fusion of dilated convolutions to capture tooth detailed information at multiple scales, and the network receptive field was adjusted. Extensive evaluation, ablation, and comparison experiments demonstrate that our method exhibits state‐of‐the‐art segmentation performance and accurate instance segmentation results, reflecting their potential applicability in clinical medicine.

During the annotation procedure of salient object detection, researchers usually locate the approximate location of the salient objects first and then process the pixels that need to be finely annotated. Following this idea, we find that the existing methods have limited exploration for solving the problem of positioning salient objects. Furthermore, no effective solution has been proposed for the hard-sample problem related to this task. Therefore, we propose dynamic scale-aware learning to learn dynamic scale weights that vary with different images to solve the first problem. Second, we design a dense sampling strategy for hard samples to construct a graph representation with samples from different classes and different confidence levels. Then, we achieve targeted feature aggregation based on the constructed graph with the help of the graph attention mechanism. We conduct extensive experiments on five benchmark datasets using comprehensive evaluation metrics. The results show that our method outperforms the current state-of-the-art approaches.