Explore the vast range of titles available.

Road extraction from high-resolution remote sensing images (HRSIs) provides critical support for downstream tasks such as autonomous driving path planning and urban planning. Although deep learning-based pixel-level segmentation methods have achieved significant progress, they still face challenges in handling occlusions caused by vegetation and shadows, and often exhibit limited model robustness and generalization capability. To address these limitations, this paper proposes the SAM2MS model, which leverages the powerful generalization capabilities of the foundational vision model, segment anything model 2 (SAM2). Firstly, an adapter-based fine-tuning strategy is employed to effectively transfer the capabilities of SAM2 to the HRSI road extraction task. Secondly, we subsequently designed a subtraction block (Sub) to process adjacent feature maps, effectively eliminating redundancy during the decoding phase. Multiple Subs are cascaded to form the multi-scale subtraction module (MSSM), which progressively refines local feature representations, thereby enhancing segmentation accuracy. During training, a training-free lossnet module is introduced, establishing a multi-level supervision strategy that encompasses background suppression, contour refinement, and region-of-interest consistency. Extensive experiments on three large-scale and challenging HRSI road datasets, including DeepGlobe, SpaceNet, and Massachusetts, demonstrate that SAM2MS significantly outperforms baseline methods across nearly all evaluation metrics. Furthermore, cross-dataset transfer experiments (from DeepGlobe to SpaceNet and Massachusetts) conducted without any additional training validate the model’s exceptional generalization capability and robustness.

Change detection plays a crucial role in remote sensing by identifying surface modifications between two sets of temporal remote sensing images. Recent advancements in deep learning techniques have yielded significant achievements in this field. However, there are still some challenges: (1) Existing change feature fusion methods often introduce redundant information. (2) The complexity of network structures leads to a large number of parameters and difficulties in model training. To overcome these challenges, this paper proposes a Multi-Scale Feature Subtraction Fusion Network (MFSF-Net). It comprises two primary modules: the Multi-scale Feature Subtraction Fusion (MFSF) module and the Feature Deep Supervision (FDS) module. MFSF enhances change features and reduces redundant pseudo-change features. FDS provides additional supervision on different scales of change features in the decoder, improving the training efficiency performance of the network. Additionally, to address the problem of imbalanced samples, the Dice loss strategy is introduced as a means to mitigate this issue. Through comprehensive experiments, MFSF-Net achieves an F1 score of 91.15% and 95.64% on LEVIR-CD and CDD benchmark datasets, respectively, outperforming six state-of-the-art algorithms. Moreover, it attains an improved balance between model complexity and performance, showcasing the efficacy of the proposed approach.

Background subtraction is a challenging and fundamental task in computer vision, which aims at segmenting moving objects from the background. Recently, the attention mechanism has become a hot topic in the neural network. The algorithms based on encoder-decoder and multi-scale type network perform impressive results in the domain of background subtraction. In this paper, we propose a multi-scale inputs and labels (MSIL) model which is based on the encoder-decoder type network and the channel attention. The multi-scale fusion encoding (MSFE) module aims to utilize multi-scale inputs effectively, which can fuse the high-level and low-level features details. The channel attention (CA) module is introduced to connect the encoder and decoder to model channel-wise attentions. The multi-label supervision decoding (MLSD) module helps to learn richer hierarchical features and achieves better performance by the new multi-label supervision. The proposed model is also evaluated on the CDnet-2014 dataset and the LASIESTA dataset, which demonstrate the effectiveness and superiority of the proposed model by an average F-Measure of 0.9851 and 0.9633, respectively. In addition, scene independent evaluation experiments on the CDnet-2014 dataset demonstrate the effectiveness of the model on unseen videos.

Accurate coronary artery segmentation from two-dimensional Digital Subtraction Angiography (DSA) images is paramount for robot-assisted percutaneous coronary intervention (PCI). Still, it is severely challenged by complex background artifacts, the intricate morphology of fine vascular branches, and frequent discontinuities in segmentation. These inherent difficulties often render conventional segmentation approaches inadequate for the stringent precision demands of surgical navigation. To address these limitations, we propose a novel deep learning framework incorporating a Composite Feature Extraction Module (CFEM) and a Multi-scale Composite Attention Module (MCAM) within a U-shaped architecture. The CFEM is meticulously designed to capture tubular vascular characteristics and adapt to diverse vessel scales. In contrast, the MCAM, strategically embedded in skip connections, synergistically integrates multi-scale convolutions, spatial attention, and lightweight channel attention to enhance the perception of fine branches and model long-range dependencies, thereby improving topological connectivity. Additionally, a combined Dice-Focal loss function is employed to optimize segmentation boundary accuracy and mitigate class imbalance jointly. Extensive experiments on the public ARCADE dataset demonstrate that our method significantly outperforms state-of-the-art approaches, achieving a Dice coefficient of 76.74%, a clDice of 50.30%, and an HD95 of 57.84 pixels. These quantitative improvements in segmentation accuracy, vascular connectivity, and edge precision underscore its substantial clinical potential for providing robust vascular structure information in robot-assisted interventional surgery.

Vessel segmentation in digital subtraction angiography (DSA) is of great significance for the diagnosis, evaluation and detection of cerebral diseases. Manual segmentation is relatively time-consuming and subjective, so that automatic cerebrovascular segmentation technology has good application value in the treatment of cerebrovascular diseases. The traditional segmentation algorithm performs poorly because of the complexity of cerebrovascular structure, large-scale changes, and the impact of noise such as artifacts in DSA. In this work, using depth learning method and attention mechanism, we propose a spatial multi-scale attention U improved network (SMAU-Net) for vessel segmentation of DSA images. The network mainly consists of three parts: multi-scale spatial attention module, feature aggregate module, and detail supervision module. Using various attention mechanisms to pay attention to scale, space and channel information, the semantic, edge information and thin vessel features are enhanced. We applied the proposed method to the benchmark retinal vessel dataset CHASE and DSAC (cerebral DSA imaging dataset made in our laboratory). The experimental results show that the proposed SMAU-Net achieves the F1 score of 86.32% and the precision of 90.04%, which is superior to other models and is an improvement of 1.98 and 5.71% over the baseline U-Net. The experiment also proves that the method can be extended to various vascular segmentation tasks and has good visual diagnosis quality.