Improved YOLOv8 Security Inspection Image for Detecting Prohibited Items

When using X-rays to detect prohibited items during security checks, detection error due to the variable target size, mutual occlusion, and image characteristics limitations. The paper presents an improved model of YOLOv8 to address the complex background of security check images, low detection accuracy of occluded targets, and complex detection models. The C2_MHSE module was designed in the neck region to more effectively extract local information while maintaining global features, enabling the extraction of target details even in complex backgrounds. PP-LCNet was adopted as the backbone network, significantly reducing the parameters and computational complexity and optimizes the positioning accuracy of prediction boxes using Inner-IoU loss function, significantly making the model more accurate in detecting threats. Analysis on the SIXray dataset highlights that the refined model upholds swift detection performance while increasing mAP50 by 1.85%, the detection speed is 111 frames per second. Compared to similar networks, it demonstrates significant advantages in both accuracy and computational speed, thus proving the effectiveness of the designed algorithm.