YOLOv7-CBAM for Similar Casting Inspection

The process of classifying industrial castings and putting them into boxes heavily relies on human vision for judgment. When castings share similar colors, shapes, and sizes, manual classification often leads to misjudgments and inefficiencies, significantly impacting economic outcomes and product reputation. Consequently, rapid, accurate, and automated inspection and identification of industrial castings have become pivotal tasks. This study focuses on the arc-shaped pipe castings primarily used in firefighting applications. Leveraging the distinguishing feature of varying identification strings on similar castings, the research employs object detection techniques from computer vision. It integrates cameras for inspection and recognition, thereby considerably reducing costs and enhancing product efficiency. The main research work is to integrate the spatial and channel attention module (CBAM), which can improve the representation and differentiation of features, into the YOLOv7 framework, to better focus on feature information, capable of enhancing feature representation and discriminability. Moreover, to enhance the generalization and robustness of the model, data enhancement is carried out. At the same time, to strengthen the feature differentiation, the study innovatively introduces a dual-label annotation approach tailored to cases where similar castings only differ in identification strings. Comparative experiments demonstrate that CBAM-enhanced YOLOv7 achieves a MAP@0.5 of 99.91%, with a frames-per-second (FPS) rate of 33.41, it fully meets the requirements of accuracy and speed of casting inspection. Exceptional on-site performance of the model in factory tests validates the proposed method's effectiveness, offering valuable insights and strategies for addressing analogous challenges in industrial environments.