With the advancement of industrial automation, an increasing number of intelligent devices are beginning to replace traditional manual labor. Particularly in the bead inspection process, the application of intelligent systems is gradually supplanting manual inspection, significantly reducing labor costs. However, current intelligent bead defect identification methods still suffer from insufficient accuracy, slow response times in selective inspection, and high rates of false detection and rejection. To meet the demand for high-speed, high-precision selective inspection of bead bubble defects in industrial production, this paper proposes a high-speed online defective bead detection and rejection system, achieving 360 \(^{\circ}\) high-speed inspection of beads. Secondly, based on the OTSU algorithm for image segmentation, the system preliminarily locates the bead outline and accurately extracts the position and grayscale information of the bead using an edge detection algorithm. Finally, an improved ResNet-based deep learning algorithm is proposed, which can analyze the extracted grayscale information in various scenarios to achieve defect classification, greatly satisfying the accuracy requirements for high-speed online selective inspection in industrial settings and effectively reducing the rates of false rejection and detection. The system’s inspection method was tested and verified through human-machine re-inspection, showing that the method achieves a selective inspection accuracy for bubble defects of up to 99.95%, a rejection rate of approximately 99.99%, a false detection rate of about 0.13%, a false rejection rate of 0.02%, and can inspect 13,400 beads per minute, meeting the efficiency requirements for online inspection. This method provides strong technical support for the high-quality inspection level in the cigarette bead industry.