Gold wire bonding is a critical interconnection technique for integrated circuit, connecting chips to substrates and enabling communication through slender wires. Gold wires are characterized by slenderness, small size, and multi-angle distribution. Meanwhile, manufacturing errors during gold wire bonding further complicate quality inspection. In this paper, we propose a fully automated gold wire bonding quality inspection framework to detect missing and misconnection wires. Given the small diameter of the gold wire (typically around 0.025 mm) and the complexity of the background, capturing high-quality images is crucial for effective quality detection. Therefore, we first design a gold wire image capture system that integrates a high-resolution camera and lens with a high-precision three-axis motion platform. This system can automatically acquire image data for both the product to be inspected and the defect-free template. Subsequently, we introduce a Slenderness Attention Module (SAM) to detect slender wires with arbitrary orientations. We employ an attention mechanism that combines channel and spatial attention, facilitating precise identification from complex background images. We then establish a Wire Quality Evaluation (WQE) model for gold wire bonding quality inspection, which integrates geometric information and data from the manufacturing process. The experimental results on our gold wire dataset demonstrate the effectiveness of the proposed method in gold wire bonding quality inspection, achieving an average precision (AP) of 85.8%, with precision (P) and recall (R) rates of 88.1% and 88.8%, respectively. The WQE model accurately identifies missing and misconnection wires. It outperforms IoU-based and similarity-based methods by detecting all defects in 16 representative images, including dense and sparse wire arrangements.

Woven composite fan blades are renowned for their superior material properties and play a crucial role in the performance of aero-engines. During the manufacturing, the uniformity of fan blade preforms is a critical factor that directly affects engine reliability. However, it’s challenging to accurately detect interlaced points due to their dense distribution and varied angular orientations caused by the weaving process and mechanical tension, making the uniformity detection of preform surfaces even more difficult. In this paper, we develop a novel method for detecting the uniformity of woven aeroengine fan blade preforms. First, we design high-resolution surface image acquisition equipment tailored to the forming and weaving characteristics of preforms. Subsequently, we propose a two-stage uniformity detection framework, termed HDID Framework. The first stage of this framework is the angular rectification network (ARNet), which standardizes the orientation of interlaced points to a horizontal alignment using a rotational object detection method. The second stage involves the dense feature refinement network (DFRNet), which utilizes multi-scale feature extraction modules and cross-attention mechanisms to effectively transform the dense detection problem into a feature point prediction task. This enables precise identification of densely interlaced points, ensuring robust performance in uniformity analysis preform detection. Ablation studies verify the effectiveness of each module, and comparative experiments demonstrate superior performance over state-of-the-art (SOTA) networks. Field measurements further validate the network’s practical applicability.

Three-dimensional (3D) single-object tracking (3D SOT) is a fundamental yet not well-solved problem in 3D vision, where the complexity of feature matching and the sparsity of point clouds pose significant challenges. To handle abrupt changes in appearance features and sparse point clouds, we propose a novel 3D SOT network, dubbed CDTracker. It leverages both cosine similarity and an attention mechanism to enhance the robustness of feature matching. By combining similarity embedding and attention assignment, CDTracker performs template and search area feature matching in a coarse-to-fine manner. Additionally, CDTracker addresses the problem of sparse point clouds, which commonly leads to inaccurate tracking. It incorporates relatively dense sampling based on the concept of point cloud segmentation to retain more target points, leading to improved localization accuracy. Extensive experiments on both the KITTI and Waymo datasets demonstrate clear improvements in CDTracker over its competitors.