Fengyi Liu

The monocular vision-based pose estimation of non-cooperative target spacecraft is vital for tasks like on-orbit servicing and debris removal. While deep learning has improved monocular spacecraft pose estimation, existing methods suffer from limitations. First, the prevailing two-stage methods separate object detection and pose estimation processes, lacking end-to-end training and involving redundant feature extraction. Second, an over-reliance on convolutional neural networks (CNNs) can result in excessive dependence on texture and inadequate long-range dependency modeling. To address these drawbacks, we propose a Deformable Transformer-based Single-stage End-to-end SpaceNet (DTSE-SpaceNet). This network dynamically fuses features from multiple scales to predict keypoints, from which pose parameters are derived using the Perspective-n-Points (PnP) method. Furthermore, a novel shape loss function improves keypoint geometric accuracy and reduces outliers and enhancing performance. Extensive experiments on multiple public benchmark datasets demonstrate competitive performance and strong generalization capability, with computation and parameter advantages over two-stage methods.

Although image recognition and classification methods based on deep neural networks (DNNs) have accomplished remarkable achievements in the domain of computer vision, more and more research in recent years has shown that these models may have vulnerabilities that can be easily exploited by attackers, leading to incorrect output results. Especially in the realm of object detection, the incorporation of meticulously crafted patches onto images can potentially deceive DNN-based detectors, resulting in erroneous output outcomes. These patches can be carefully designed to greatly interfere with and mislead the output results of deep neural networks. These misleading output results may lead to serious security issues and losses, so it is becoming increasingly important to study and strengthen the robustness and security of DNNs based target detectors. However, previous methods did not consider the interpretability of the model, and this paper proposes a new mothed to adversarial patch attacks. First, we employ the model interpretable algorithm Grad-CAM to obtain the patch's location within the model's region of interest for the target. Subsequently, we calculate the precise patch placement and apply it accordingly. Following that, we utilize a loss function to optimize patch, ensuring its effectiveness. The experimental results obtained from the DOTA dataset indicate that our proposed method effectively degrades the performance of the target detection model. Secondly, we conducted experiments to improve the transferability of adversarial patches across different models, and verified that using patches to cover the regions of interest in the model improves the transferability of patches.

Structural health monitoring (SHM) technology is of great importance to ensure the long-term and reliable operation of engineering structures. The monitoring systems for large structures are characterized by large numbers of measurement points and complex signal. In recent years, deep learning has been widely used in many fields, such as intelligent fault diagnosis and health monitoring due to its excellent feature extraction and nonlinear fitting ability. However, the existing deep networks for condition monitoring have some drawbacks. First, the multisensor information fusion is rarely considered, which is quite common in SHM systems. Second, most of the proposed models by current research are with large parameters and high complexity. To solve the problems mentioned above, we propose a lightweight SHM framework suitable for the multisensor monitoring tasks. In other words, the adaptive multisensor fusion network (ASF-Net) is proposed to dynamically mix with multichannel input samples and extract the efficient feature on different scales. Moreover, a joint loss function is established for network compression by transferring the knowledge from bulky teacher network to lightweight student network. Finally, to reduce the performance degradation during model compression, a multigeneration knowledge distillation (MGKD) strategy is proposed. The experiments on two SHM datasets show that the proposed method can achieve excellent classification accuracy, and model performance degradation of less than 2% can be obtained with at most 40.24 times less parameters and 128.82 times less computation, which has a great prospect on resource-constrained platforms.