In frequency-modulated continuous-wave (FMCW) radar interference suppression based on signal reconstruction, the pruned exact linear time (PELT) algorithm is used to detect the time positions of the interference. Due to the uncertain penalty factor of the PELT algorithm, the exactness of the position detection is reduced; thus, the suppression performance is degraded. We propose a PELT algorithm with a known change number (PELT-KCN), where a known change number is used to calculate the optimal penalty factor such that the high accuracy of the algorithm can be guaranteed. After interference recognition, the beat signal is separated into two parts: the undamaged signal and the damaged signal. The former is utilized to restore the latter through an autoregressive (AR) model. In simulations and field experiments, we applied our proposed PELT-KCN algorithm to the interference suppression method and verified its performance. Our method can accurately detect the time positions of interference and effectively improve the signal-tonoise ratio (SNR) of the detected targets.
A magnetic ground state is revealed for the first time in zigzag-edged carbon nanoscrolls (ZCNSs) from spinunrestricted density functional theory calculations. Unlike their flat counterpart—zigzag-edged carbon nanoribbons, which are semiconductors with spin-degenerate electronic structure—ZCNSs show a variety of magnetic configurations, namely spin-selective semiconductors, metals, semimetals, quasi-half-metals, and half-metals. To the best of our knowledge, this is the first discovery of quasi-half-metals and half-metals in a pure hydrocarbon without resort to an external electric field. In addition, we calculated the spin-dependent transportation of the semiconducting ZCNSs with 12 and 20 zigzag chains, and found that they are 13% and 17% at the Fermi level, respectively, suggesting that ZCNS can be an effective spin filter.
Deep convolutional neural networks (CNNs) are of great improvement for single image super‐resolution (SISR). However, most existing SISR pre‐trained models can only perform single image restoration and the upscale factors cannot be non‐integers, which limits its application in real‐world scenarios. In this letter, an enhanced dual branches network (EDBNet) in upsampling network is proposed to generate arbitrary‐scale super‐resolution (SR) images. Specifically, the authors design a scale‐guidance upsampling module (SGU) by adding the scale factors and pixel‐level features to guide the weights of convolution. The SGU module performs discriminant learning for each instance in the same batch. Extensive experiments on four benchmark datasets show that the proposed method can achieve superior SR results.
9 September 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Road information extracted from high-resolution remote sensing images is crucial for urban planning and traffic navigation. However, challenges such as complex and variable geometric and contextual features of roads, occlusion by trees and buildings, and labeling errors in datasets present significant difficulties. To address these issues, we propose the UGD-DLinkNet model. Specifically, attention mechanisms are employed to enhance road feature extraction in occluded and complex regions by adaptively focusing on informative spatial and channel-wise cues, while model uncertainty estimation helps reduce the impact of noisy annotations by guiding learning toward more reliable predictions. First, we introduce a hybrid attention module to strengthen the encoder's ability to capture essential features and structural information. Second, a self-attention unit is integrated into the bridging network to form a dilated convolution attention module (DCAM), improving the perception of semantic features across scales. Third, a channel attention module refines skip connections, bridging shallow and deep semantic information. Finally, we incorporate Monte Carlo dropout (MC dropout) and propose an uncertainty-guided knowledge distillation strategy to mitigate labeling noise. Experiments conducted on the Massachusetts, CHN6-CUG, and DeepGlobe benchmark datasets demonstrate the comprehensive superiority of UGD-DLinkNet. On the Massachusetts dataset, our model achieves the highest F1-score of $78.75\%$. On CHN6-CUG, it obtains the best overall performance with an F1-score of $76.50\%$ and an IoU of $61.95\%$, exceeding the best results among recent mainstream methods by $4.68\%$ and $2.33\%$, respectively. On DeepGlobe, UGD-DLinkNet also achieves top performance, with an F1-score of $81.49\%$ and an IoU of $68.77\%$, highlighting its strong generalization and robustness across diverse scenarios. The code will be available at https://github.com/VisionENet/UGD-DLinkNet.
30 August 2023 in IEEE Transactions on Circuits and Systems II: Express Briefs
Contrastive representation learning purposes to learn rich point cloud representations from unlabeled data. However, it currently lacks well-designed data augmentation, which can lead to poor regularization and generalization in the point cloud domain. Mixup has been shown to be a simple and effective data augmentation in the image domain. Yet, it was found ineffective to directly applying it to point cloud domain. In this brief, we propose a novel mixup strategy based on the cosine similarity algorithm in point cloud contrastive representation learning. First, we adaptively select pairs of samples with high dissimilarity in the mini-batch for mixing. The Principal Component Analysis (PCA) dimensionality reduction algorithm is used to reduce the cost when calculating the similarity. Then we introduce N-pair contrastive loss that fits to the virtual labels assigned for mixing. Our method outperforms other methods in a diverse range of downstream tasks including 3D object classification and few-shot object classification. Our method can be extended to manifold mixup for better regularization. Meanwhile, the ablation experiments demonstrate that our method is an effective data augmentation strategy.
15 January 2024 in IEEE Transactions on Circuits and Systems II: Express Briefs
Depth estimation from a single image is a challenging yet crucial task for reconstructing 3D structures and inferring scene geometry. However, most methods struggle to effectively leverage global information at high resolutions to establish long-distance dependencies between distant pixels, thereby affecting the accurate prediction of depth for objects in shadows and faraway regions. To address this issue, we propose GlobalDepth, which utilizes the Global Information Aggregation Module (GIAM) based on self-attention mechanism to model long-distance dependencies in the spatial domain. We introduce the Global Information Enhancement Module (GIEM) based on multi-scale global average pooling to enhance global information in depth feature maps. Experimental results on the NYU-Depth-v2, KITTI, and SUNRGB-D datasets demonstrate that GlobalDepth outperforms existing methods in terms of performance and generalization capability.