Real-time detection of tunnel-lining cracks is hindered by low contrast, irregular morphology, and complex background textures. We propose ADRL-DETR (Adaptive Dual-Path Representation Learning DETR), a lightweight transformer-based detector that introduces a spatial–scale dual-path adaptive mechanism to enhance spatial feature discrimination and enable cross-scale fusion guided by dynamic receptive-field modulation. The Global–Local Adaptive Filtering Block (GLAFB) serves as the multi-spatial feature expression path, adaptively separating cracks from background via complementary global–local filtering, while the Scale-Adaptive Fusion Block (SAFB) functions as the cross-scale fusion path, integrating multi-dilated convolutions and channel attention to achieve dynamic receptive-field adjustment and synergistic feature optimization. On the MixCrack-6K dataset, ADRL-DETR attains 87.8% Precision and 78.8% AP50with 51.0 GFLOPs, outperforming the baseline by 3.2% Precision and 3.1% AP50. The proposed framework achieves an effective trade-off between accuracy and efficiency, offering an interpretable and deployable solution for real-time tunnel inspection.

Hyperspectral image denoising is crucial for accurate extraction of spectral information. However, current convolutional neural network (CNN)-based methods have inherent limitations, while Transformer- based methods suffer from high computational complexity when processing global contextual information. To address this problem, we designed a hybrid Mamba-CNN context interaction module and constructed a U-shaped hierarchical encoder-decoder network (MUNet). The network takes the pixel-scale as input to maximize the preservation of image information and employs state-space model (SSM)-based Mamba blocks to efficiently capture global semantic information, while using convolution to extract local features. This enhances the modeling of global and local features for better denoising. Extensive experiments on synthetic and real hyperspectral image (HSI) datasets showed that the proposed MUNet achieves better performance than other state-of-the-art techniques.

Remote sensing images are different from ordinary images in that they have higher resolution, contain information of a larger area, and are characterized by strip-like objects in many scenes. The traditional Transformer model based on the moving window to calculate the attention is difficult to obtain the overall features when extracting the features of strip-shaped objects and is easily interfered by the surrounding features. To address this problem, this paper innovatively designs an axial Transformer module and constructs a U-shaped hierarchical encoder-decoder structure network (ATUNet). The network improves its ability to extract global features and resist interference from irrelevant features through the axial attention mechanism. We synthesize multiple test sets with noise levels for experiments using three datasets, NWPU-RESISC45, UCMerced_LandUse, and OPTIMAL-31. The experiments show that our network has good resistance to high noise and generalization ability.

A new quaternion convolutional attention denoising network known as DeQCANet for color random-valued impulse noise removal is proposed in this paper. First, a structural information extraction block based on a dilated convolution operation is designed to extract the structure and detail feature map. Subsequently, a quaternion convolutional neural network with a new quaternion map construction strategy is implemented to gain the color features across channels further. Finally, to integrate the global and local features, a feature enhancement block based on the attention mechanism is proposed to guide the network for impulse noise denoising. The experimental results demonstrate that the proposed denoising technique exhibits competitive performance compared to other well-known color image denoising methods.

Despite numerous progresses in the past decades, 3D shape acquisition techniques remain a threshold for various 3D face based applications. Moreover, advanced 2D data generative models based on the deep networks may not be directly applicable for 3D objects. In this work, we propose a geometry sampling approach to bridge the gap between unstructured 3D face models and the powerful deep networks towards an unsupervised 3D face generative model. Specifically, we devise a geometry sampling approach to obtain a structured representation of 3D faces, which enable us to adapt the 3D faces to the Deep Convolution Generative Adversarial Network (DCGAN) for 3D face generation. We have demonstrated the effectiveness of our generative model by producing a large variety of 3D faces with different facial expressions.

Impulse noise detection is important to the restoration of color images contaminated by impulse noise in switching vector median filters. To increase detection accuracy, an effective color-impulse detector is presented. A new color distance metric based on quaternion theory is proposed. The proposed color distance metric is used to calculate the local density of a color pixel. A hard thresholding strategy is used to determine whether a color pixel is corrupted by impulse noise or not (i.e., an outlier). The noisy pixels detected will be restored by a weighted vector median filter, while the noise-free pixels remain unchanged. The experimental comparisons show that the proposed algorithm can obtain lower false and miss detection rate, and produces better performance in terms of peak signal-to-noise ratio and feature similarity measures, compared to other well-known color image filtering methods.