The perception and processing of transparent objects face significant challenges in various applications, primarily due to the limitations of traditional sensors. These sensors often struggle to capture the complete depth information of transparent objects, mainly due to the refraction and reflection of light on their surfaces, as well as the lack of visible texture. Previous research has explored the use of deep learning models to generate complete depth maps from RGB images and corrupted depth data obtained from depth sensors. However, existing methods suffer from design flaws that limit the effectiveness of depth completion. To address this issue, we propose TDCNet, a novel dual-branch CNN-Transformer parallel network specifically designed for depth completion of transparent objects. Our framework consists of two distinct branches: one focuses on feature extraction from partial depth maps, while the other processes RGB-D images. Experimental results demonstrate that our model achieves state-of-the-art performance on multiple public datasets. The code and pre-trained models are publicly available at https://github.com/XianghuiFan/TDCNet.

The simple lens computational imaging method provides an alternative way to achieve high-quality photography. It simplifies the design of the optical-front-end to a single-convex-lens and delivers the correction of optical aberration to a dedicated computational restoring algorithm. Traditional single-convex-lens image restoration is based on optimization theory, which has some shortcomings in efficiency and efficacy. In this paper, we propose a novel Recursive Residual Groups network under Generative Adversarial Network framework (RRG-GAN) to generate a clear image from the aberrations-degraded blurry image. The RRG-GAN network includes dual attention module, selective kernel network module, and residual resizing module to make it more suitable for the non-uniform deblurring task. To validate the evaluation algorithm, we collect sharp/aberration-degraded datasets by CODE V simulation. To test the practical application performance, we built a display-capture lab setup and reconstruct a manual registering dataset. Relevant experimental comparisons and actual tests verify the effectiveness of our proposed method.

This study proposes a deep metric learning-based pavement distress classification method to address critical limitations in conventional approaches, including their dependency on large training datasets and inability to incrementally learn new categories. To resolve high intra-class variance and low inter-class distinction in distress images, we design a CNN head with multi-cluster centroins trained via SoftTriple loss, simultaneously maximizing inter-class separation while establishing multiple intra-class centers. An adaptive weighting strategy combining sample similarity and class priors mitigates data imbalance, while soft-label techniques reduce labeling noise by evaluating similarity against support-set exemplars. Evaluations on the UAV-PDD2023 dataset demonstrate superior performance—3.2% higher macro-recall than supervised learning, and 6.7%/8.5% improvements in macro-F1/weighted-F1 over iCaRL incremental learning—validating the method’s effectiveness for real-world road inspection scenarios with evolving distress types and limited annotation.

Space-based on-orbit servicing missions impose stringent requirements for precise identification and localization of satellite components, while existing detection algorithms face dual challenges of insufficient accuracy and excessive computational resource consumption. This paper proposes SLD-YOLO, a lightweight satellite component detection model based on improved YOLO11, balancing accuracy and efficiency through structural optimization and lightweight design. First, we design RLNet, a lightweight backbone network that employs reparameterization mechanisms and hierarchical feature fusion strategies to reduce model complexity by 19.72% while maintaining detection accuracy. Second, we propose the CSP-HSF multi-scale feature fusion module, used in conjunction with PSConv downsampling, to effectively improve the model’s perception of multi-scale objects. Finally, we introduce SimAM, a parameter-free attention mechanism in the detection head to further improve feature representation capability. Experiments on the UESD dataset demonstrate that SLD-YOLO achieves measurable improvements compared to the baseline YOLO11s model across five satellite component detection categories: mAP50 increases by 2.22% to 87.44%, mAP50:95 improves by 1.72% to 63.25%, while computational complexity decreases by 19.72%, parameter count reduces by 25.93%, model file size compresses by 24.59%, and inference speed reaches 90.4 FPS. Validation experiments on the UESD_edition2 dataset further confirm the model’s robustness. This research provides an effective solution for target detection tasks in resource-constrained space environments, demonstrating practical engineering application value.