Detection of small rod-end joint bearings via deep feature fusion and confidence propagation clustering

• Jinming Pen ^[1] ; Ruifeng Ye ^[2] ; Song Lan ^[3] ; Tenghao Xiao ^[1] ; Chen Xu ^[1] ; Yancong Song ^[1]
  1. [1] Fujian University of Technology

     Fujian University of Technology

     China

     
  2. [2] College of Artificial Intelligence, Yango University, Fuzhou, 350015, China
  3. [3] College of Automation Engineering, Fujian College of Water Conservancy and Electric Power, Yong an, 366000, China

  Mostrar afiliaciones +
• Localización: Métodos numéricos para cálculo y diseño en ingeniería: Revista internacional, ISSN 0213-1315, Vol. 40, Nº 3, 2024
• Idioma: inglés
• DOI: 10.23967/j.rimni.2024.10.56511
• Enlaces
  • Texto completo
• Resumen

  • Machine vision is used to detect dense, small rod-end joint bearings in sliding ball surfaces with little feature information and high variability. However, this leads to inaccurate identification, affecting production efficiency. This study proposes a deeplearning object-detection algorithm model that allows the network to retain more semantic information. We introduced the space-to-depth convolution (SPD-Conv) step-free convolution module to improve the backbone network and developed a multi-level feature fused SPD (MFSPD) deep feature fusion module to redesign the neck network to improve the feature extraction ability and detection accuracy for small targets. Furthermore, we added a small P4 detection head in the head network (i.e., prior box acquisition on the dataset using the weighted k-means algorithm), increased the matching degree of the prior box and feature layer, and accelerated the model convergence. To improve the confidence propagation clustering (CP-Cluster) analysis algorithm for post-processing, we optimized the prediction box confidence degree and detection speed. The algorithm performance was evaluated on homemade, T-LESS, and COCO datasets. The mAP@.5 values of the target detection algorithm for the homemade and T-LESS datasets were 96.9% and 93.8%, respectively, and the mAP was 55.9% for the COCO dataset. The experimental results indicate that the algorithm has a high detection accuracy and good feature extraction ability. Thus, it has considerable advantages for small-object detection and provides a reference for the detection of small parts.