Xiangsheng Huang

Due to the complexity of molecules, molecular learning requires a large amount of molecular data. However, labeled data is typically limited, making self-supervised pretraining methods essential. Despite this, current pretraining methods often fail to sufficiently focus on both local and global molecular information. In this study, we propose a multi-modality self-supervised learning framework that simultaneously captures local and global information. Specifically, we encode SMILES sequences and molecular graphs separately and use a unified fusion approach to strengthen the interaction between the two modalities. Moreover, in the molecular graph encoding, we independently capture global and local information, and enhance the attention to bond features through information fusion. Additionally, we introduce the FA-FFN module to aggregate periodic features of the molecule. Experimental results show that MoleTGL exhibits superior performance compared to existing methods on seven classification tasks and six regression tasks related to molecular property prediction, and ablation studies confirm the effectiveness of local and global feature fusion and the superiority of the methods for acquiring local and global information.

Semantic segmentation is a fundamental task in computer vision. In multi-class scenarios, the abundance of categories, feature similarity across classes, class imbalance, and the complexity of feature space modeling pose significant challenges in delineating features and class boundaries.Recent multi-class segmentation methods often suffer from issues such as boundary ambiguity, loss of detail, and inadequate preservation of contextual information.To address these challenges, this paper introduces the Multi-scale Frequency Feature Fusion Network (MFN), a convolutional neural network designed to effectively integrate multi-scale information. First, the Deeplab network architecture is enhanced by refining the ASPP module. Finer-grained atrous convolutions are introduced in its channels to capture multi-scale semantic information, enabling the extraction of richer and more detailed features.The network incorporates a hybrid attention module to enhance the importance of specific spatial points and strengthen feature dependencies between channels.During the feature fusion stage, the FreqFusion module adaptively smooths high-level features to address class inconsistency and boundary misalignment effectively.And a combined loss function, integrating cross-entropy loss and the Dice coefficient, is designed to improve model training and achieve optimal segmentation performance.Extensive experiments conducted on the Cityscapes and Pascal VOC datasets demonstrate that the proposed method achieves superior performance.

Accurate detection of extremely small targets in aerial high-resolution images is critical for military and civilian applications. However, challenges such as small target size, complex backgrounds, and target deformation hinder optimal performance. We propose QLDNet, a lightweight network that addresses these issues through a synergistic combination of non-strided convolution and decoupled large-kernel convolutional attention. Specifically, the non-strided convolution constructs a quadruple aggregate and connection feature extractor (QACFE) module, which maps features into the channel dimension while preserving spatial details. The decoupled large kernel convolutional attention then leverages these channel features to effectively extract structural and edge-related low-frequency information, while simultaneously reducing computational costs and model size caused by the increased channel dimensionality after QACFE. A learnable offset mechanism is introduced, transforming the detection head into a deformable detection head. Additionally, the network incorporates a PAFPN (Path Aggregation Feature Pyramid Network) structure to efficiently extract multi-scale features. During inference, a tailored approach performs pixel-level multi-scale detection. This enhances small target detection by merging features and pixels through dual-scale fusion, integrating multi-scale feature extraction with multilevel feature integration.Experimental results demonstrate that QLDNet achieves efficient and accurate small target detection, with an accuracy rate as high as 94.0\% and an inference time of 0.12 seconds per image, meeting the system's real-time requirements with fewer parameters and satisfying processing speeds.The code can be find at https://github.com/Sjl185721/QLDNet-v1.git. Mathematics Subject Classification (2020) MSC68T05 · MSC 86Axx · MSC 68T10

Accurate semantic segmentation and automatic thickness measurement of the glomerular basement membrane (GBM) can aid pathologists in carrying out subsequent pathological diagnoses. The GBM has a complex ultrastructure and irregular shape, which makes it difficult to segment accurately. We found that the shape of the GBM is striped, so we proposed an RSP model to extract both the strip and square features of the GBM. Additionally, grayscale images of the GBM are similar to those of surrounding tissues, and the contrast is low. We added an edge attention mechanism to further improve the quality of segmentation. Moreover, we revised the pixel-level loss function to consider the tissues around the GBM and locate the GBM as a doctor would, i.e., by using the tissues as the reference object. Ablation experiments with each module showed that SSPNet can better segment the GBM. The proposed method was also compared with the existing medical semantic segmentation model. The experimental results showed that the proposed method can obtain high-precision segmentation results for the GBM and completely segment the target. Finally, the thickness of the GBM was calculated using a skeleton extraction method to provide quantitative data for expert diagnosis.

Most two-stage detection and tracking methods currently adopt IoU distance and appearance features when matching target detection results with trajectory predictions. However, this approach has shortcomings in two main aspects: 1. The selection of feature extraction networks often relies on established methods, such as Re-ID-related methods, which are typically complex and significantly slow down algorithm processing speed. 2. Feature extraction networks lack consistency when extracting single-frame features, leading to difficulties in matching incomplete objects with previous features when occlusions occur, resulting in incorrect updates of target identities. To address these limitations, this study introduces a lightweight feature extraction network that combines current target features with preceding information for fusion. This network enhances data processing speed while obtaining smooth object features with temporal coherence. Moreover, considering variations in feature granularity among different targets, various feature matching approaches are employed to handle detection boxes under different conditions, maximizing information efficiency and conserving computational resources. An adaptive multi-feature extraction method is also proposed to accommodate diverse requirements across different scenarios. The matching strategy developed allows the tracker to serve as a universal and practical solution for various tracking scenarios within a unified framework, eliminating the need for manual parameter tuning. The effectiveness of PreAFTrack is demonstrated through comprehensive evaluations on three major MOT datasets: MOT17, MOT20, and DanceTrack.