The existing segmentation-based scene text detection methods mostly need complicated post-processing, and the post-processing operation is separated from the training process, which greatly reduces the detection performance. The previous method, DBNet, successfully simplified post-processing and integrated post-processing into a segmentation network. However, the training process of the model took a long time for 1200 epochs and the sensitivity to texts of various scales was lacking, leading to some text instances being missed. Considering the above two problems, we design the text detection Network with Binarization of Hyperbolic Tangent (HTBNet). First of all, we propose the Binarization of Hyperbolic Tangent (HTB), optimized along with which the segmentation network can expedite the initial convergent speed by reducing the number of epochs from 1200 to 600. Because features of different channels in the same scale feature map focus on the information of different regions in the image, to better represent the important features of all objects in the image, we devise the Multi-Scale Channel Attention (MSCA). Meanwhile, considering that multi-scale objects in the image cannot be simultaneously detected, we propose a novel module named Fused Module with Channel and Spatial (FMCS), which can fuse the multi-scale feature maps from channel and spatial dimensions. Finally, we adopt cross-entropy as the loss function, which measures the difference between predicted values and ground truths. The experimental results show that HTBNet, compared with lightweight models, has achieved competitive performance and speed on Total-Text (F-measure:86.0%, FPS:30) and MSRA-TD500 (F-measure:87.5%, FPS:30).

Non-Euclidean data, such as social networks and citation relationships between documents, have node and structural information. The Graph Convolutional Network (GCN) can automatically learn node features and association information between nodes. The core ideology of the Graph Convolutional Network is to aggregate node information by using edge information, thereby generating a new node feature. In updating node features, there are two core influencing factors. One is the number of neighboring nodes of the central node; the other is the contribution of the neighboring nodes to the central node. Due to the previous GCN methods not simultaneously considering the numbers and different contributions of neighboring nodes to the central node, we design the adaptive attention mechanism (AAM). To further enhance the representational capability of the model, we utilize Multi-Head Graph Convolution (MHGC). Finally, we adopt the cross-entropy (CE) loss function to describe the difference between the predicted results of node categories and the ground truth (GT). Combined with backpropagation, this ultimately achieves accurate node classification. Based on the AAM, MHGC, and CE, we contrive the novel Graph Adaptive Attention Network (GAAN). The experiments show that classification accuracy achieves outstanding performances on Cora, Citeseer, and Pubmed datasets.

The safety and reliability of urban intelligent transportation systems(ITS) largely depend on accurate and sufficient scene perception, especially detecting and understanding text information such as traffic signs and road instructions. Urban scene text has arbitrary shape boundaries and overlapped instances, which is hard to represent in the spatial domain. To solve the above challenges, we contrive Discrete Cosine Transform Network(DCTNet), in which the discrete cosine transform boundary representation method is proposed to convert the scene text boundary in the spatial domain to the frequency domain, and then generate the exact scene text boundary through the inverse discrete cosine transformation, thus solving the geometric representation and overlapping problems of arbitrary shape boundaries. The backbone is designed based on Contrastive Language-Image Pre-Training(CLIP) to extract refined features. At the same time, the combined Deformable Convolutional Networks(DCN) in the detection head of DCTNet adaptively perceives categories and edge features. DCTNet has achieved SOTA performances on Total-Text and MSRA-TD500, corresponding to F1 of 90.8% and 89.6%, respectively. The high efficiency and accuracy of DCTNet provide a solid foundation for the scene perception of the intelligent transportation systems, and further improve the safety and reliability of ITS.

The existing segmentation-based scene text detection methods mostly need complicated post-processing, and the post-processing operation is separated from the training process, which greatly reduces the detection performance. The previous method, DBNet successfully simplified post-processing and integrated the post-processing into a segmentation network. However, the training process of the model took a long time for 1200 epochs and the sensitivity to texts of various scales was lacking, leading to some text instances being missed. Considering the above two problems, we design the text detection Network with Binarization of Hyperbolic Tangent(HTBNet). First of all, we propose Binarization of Hyperbolic Tangent (HTB), optimized along with which, the segmentation network can expedite the initial convergent speed by reducing the amount of epochs from 1200 to 600. Because features of different channels in the same scale feature map focus on the information of different regions in the image, to better represent the important features of all objects in the image, we devise the Multi-Scale Channel Attention(MSCA). Meanwhile considering that multi-scale objects in the image cannot be simultaneously detected, we propose a novel module named Fused Module with Channel and Spatial(FMCS), which can fuse the multi-scale feature maps from channel and spatial dimension. Finally we adopt cross entropy as the loss function, which measures the difference between predicted values and ground truths. The experimental results show that HTBNet compared with lightweight models has achieved competitive performance and speed on Total-Text(F-measure:86.0%, FPS:30) and MSRA-TD500 (F-measure:87.5%, FPS:30).

：Non-Euclidean data, such as social networks, and citation relationships between documents, has node information and structural information. Graph Convolutional Network(GCN) can automatically learn node features and association information between nodes. The core ideology of the graph convolutional network is to aggregate node information by using edge information, thereby generating a new node feature. In the process of updating node features, there are two core influencing factors. One is the number of neighboring nodes of the central node, the other is the contribution of the neighboring nodes to the central node. Due to the previous GCN methods not simultaneously considering the numbers and different contributions of neighboring nodes to the central node, we design the Adaptive Attention Mechanism(AAM). To further enhance the representational capability of the model, we utilize Multi-Head Graph Convolution(MHGC). Finally, we adopt cross-entropy(CE) loss function to describe the difference between the predicted results of node categories and the ground truth (GT). Combined with backpropagation, this ultimately achieves accurate node classification. Based on AAM, MHGC and CE, we contrive the novel Graph Adaptive Attention Network (GAAN). Experiments show that the classification accuracy has achieved outstanding performances on Cora, Citeseer and Pubmed datasets.

High-resolution synthetic aperture radar (SAR) ship detection plays a pivotal role in maritime surveillance and ocean monitoring. However, it remains challenging in practice because of the single-channel imaging modality, severe multiplicative speckle noise, and the pronounced scale imbalance in which sparse large vessels are easily under-optimized compared to the dominant small and medium instances. In this paper, we propose a Texture Perception Convolution Network (TPCNet), a practical and reproducible detection framework to improve feature extraction robustness and high-IoU localization under a unified strict-COCO evaluation protocol. TPCNet begins with a lightweight texture perception convolution (TPConv) that augments the raw SAR intensity with a local fluctuation cue to stabilize early feature representations for SAR images affected by strong multiplicative speckle noise (speckle-rich imagery). To address scale skew during training without modifying dataset splits, a region-balanced sampler (RBS) is introduced to increase the sampling probability of images, thereby improving the effective exposure of informative large-target structures. A background similarity augmentation (BSA) is proposed to enrich medium and large instances while reducing unrealistic boundary artifacts via compatible background selection and soft blending. Beyond component-level designs, the high-IoU localization is highly sensitive to geometric perturbations. Accordingly, TPCNet adopts a two-stage localization-oriented training strategy that first learns robust multi-scale representations and then refines box regression by tightening translation and scale ranges during fine-tuning. Under strict-COCO settings, TPCNet achieves SOTA performance with an AP of 71.20% on HRSID and an AP of 72.7% on SSDD. Comprehensive ablation studies demonstrate that TPConv, RBS, BSA, and the proposed finetuning strategy contribute complementary gains, providing a transparent baseline and a strong recipe for future SAR ship detection research.