6 June 2023 in IEEE Access
Recent advances in deep learning have enabled state-of-the-art performance in detecting medium and large-size objects. However, small object detection remains challenging primarily due to the scarcity of information. This paper proposes an end-to-end fusion network that integrates deep and hand-crafted features to address this limitation. A fusion module based on semantic context information is designed to enhance feature discrimination ability. Additionally, we introduce a kind of feature-contrast loss to incorporate prior knowledge into the learning of deep feature according to contrastive learning. Experiments on the MS COCO dataset (34.4% AP S ) demonstrate that our approach achieves improved detection accuracy over previous methods, especially for small objects.
7 November 2023 in IEEE Access
Tracking multiple states of the target simultaneously is currently a research hotspot in object tracking. Existing methods obtain the initial bounding box with multi-scale search, anchor-based regression, or anchor-free regression. Then, each pixel within bounding box is classified and the mask of target is used to fit a fine bounding box for accurate tracking. However, their computation complexity is high and the accuracy of mask and fine bounding box is limited by initial bounding box. Based on SiamMask, the Siamese contour segmentation network (SiamCS) is proposed for multi-state object tracking to address these issues. This end-to-end network eliminates need to pre-define anchor based on prior knowledge, and reduces hyperparameters. With SiamCS, multi-state object tracking method formulates object tracking as region classification and contour regression to obtain bounding box, contour, and mask of target at the same time. Moreover, according to geometric meaning of definite integral and difficulty of sample, difficulty sensitive contour-intersection over union loss function is proposed to solve the problem of independent regression of contour parameters. Extensive experiments on OTB100 (92.5%Precision), UAV123 (83.5%Precision), LaSOT (64.8%AUC), TrackingNet (81.5%AUC), GOT-10K (66.3%AO), and VOT2020 (53.5%EAO) show that SiamCS outperforms many state-of-the-art trackers and achieves leading performance.
16 September 2025 in Recent Advances in Electrical & Electronic Engineering (Formerly Recent Patents on Electrical & Electronic Engineering)
Introduction: Monitoring the furnace hearth activity is of great significance for maintaining the production efficiency of steel. In order to solve the problem that it is difficult to collect sufficient blast furnace process parameters in actual production, resulting in the model not being fully trained, this research proposed a method based on distance-angle metric learning to predict hearth activity in scenarios with few samples. Methods: In this research, samples were randomly selected and grouped in pairs. The proposed method extracted features through two symmetrical back propagation neural networks to get multiple pairs of feature vectors, then predicted the state of furnace hearth activity by measuring the Manhattan distance and cosine similarity between each pair of feature vectors. Results: The experimental results showed that the proposed method can extract sample features stably, it still has a good analysis effect in scenarios with few samples. The optimal accuracy rate was 99.18%, which can meet the monitoring of hearth activity in actual production. Discussion: The measurement based on distance and angle can reflect the characteristics of each category more accurately. It also achieved good performance in scenarios with few samples and provided a theoretical basis for the reasonable regulation of the hearth activity. Conclusion: Compared with the related methods, the proposed method based on distance-angle metric learning in this research has better performance. It can meet the application requirements in actual production scenarios and has great research value.
22 November 2025 in arXiv
RGB-Thermal (RGBT) tracking aims to exploit visible and thermal infrared modalities for robust all-weather object tracking. However, existing RGBT trackers struggle to resolve modality discrepancies, which poses great challenges for robust feature representation. This limitation hinders effective cross-modal information propagation and fusion, which significantly reduces the tracking accuracy. To address this limitation, we propose a novel Contextual Aggregation with Deformable Alignment framework called CADTrack for RGBT Tracking. To be specific, we first deploy the Mamba-based Feature Interaction (MFI) that establishes efficient feature interaction via state space models. This interaction module can operate with linear complexity, reducing computational cost and improving feature discrimination. Then, we propose the Contextual Aggregation Module (CAM) that dynamically activates backbone layers through sparse gating based on the Mixture-of-Experts (MoE). This module can encode complementary contextual information from cross-layer features. Finally, we propose the Deformable Alignment Module (DAM) to integrate deformable sampling and temporal propagation, mitigating spatial misalignment and localization drift. With the above components, our CADTrack achieves robust and accurate tracking in complex scenarios. Extensive experiments on five RGBT tracking benchmarks verify the effectiveness of our proposed method. The source code is released at https://github.com/IdolLab/CADTrack.
10 October 2024 in IEEE Access
Overcoming the high cost of self-attention operation in Transformer-based object detection methods and improving the detection accuracy of small objects is one of the difficulties in the field of object detection research. This paper designs a progressive assisted object detection method PaoDet based on Transformer, which uses common feature extraction backbone such as Resnet and ViT to extract multi-scale features of the input image, and uses RPN(Region Proposal Network) to extract proposals of different scales; Subsequently, a progressive modeling approach was adopted to perform self-attention and cross-attention operations on proposals of different scales from large to small, achieving feature interaction between instances, ensuring high detection efficiency and low computational complexity. During the training process, each layer of the network has certain generalization ability for detecting adjacent scale objects under the supervision of a dynamic scale division method. Compared with state-of-the-art object detection methods on COCO and UAVDT datasets, the effectiveness and superiority of the proposed method were demonstrated.