Improved X-Ray Image Contraband Detection Algorithm for YOLOv8n

To address the challenges posed by complex backgrounds, the imbalance between positive and negative samples, and the omission of small target contraband in X-ray contraband detection, an enhanced YOLOv8n algorithm is proposed. This algorithm introduces deformable convolution in place of standard convolution within the backbone feature extraction network. The deformable convolution adaptively learns features of contraband of varying sizes and shapes, thereby improving detection accuracy in complex and dynamic environments. To enhance multi-scale feature fusion and significantly improve the model's detection accuracy, the GSConv and VOV-GSCSP modules are integrated into the neck feature fusion network. Additionally, Focal DIOU Loss is employed as the boundary regression loss function to accelerate convergence and enhance the network's regression detection accuracy. Experimental results demonstrate that the proposed method achieves an accuracy of 94.8% and a [email protected] of 93.3% on the SIXray dataset, representing improvements of 5.1% and 3.1%, respectively, over the original YOLOv8n. Furthermore, the detection speed reaches 105.68 FPS, ensuring real-time performance alongside high detection accuracy.