Improved YOLOv8 based on DCNv3 and DFES for x-ray prohibited items detection

Due to the diverse placement of objects, varying sizes of luggage, and significant occlusion issues in real-world scenarios, existing x-ray detection methods for prohibited items generally suffer from inadequate detection accuracy for small target objects, which leads to missed detection and false positives. Based on this, we propose an improved YOLOv8 model (DF-YOLOv8), which incorporates deformable convolutions and deep feature extraction structure. Firstly, in the backbone of the model, we modify the standard convolutions at the front of the C2f module to deformable convolutions networks v3(DCNv3). Secondly, the deep feature extraction structure is added into deeper layers of YOLOv8 to extract more accurate global feature information. The experimental results on HIXray dataset show that DFYOLOv8 outperforms the State of The Arts (SOTA) detectors and the detection accuracy of it is as high as 85.6%, while also reducing the Floating Point Operations (FLOPs) to make the model lighter.