Research And Practice On Lightweight Contraband Detection Technology Based On Improved YOLOv5

The detection of contraband is an important measure to ensure the safety of people and transportation,but the domestic security inspection method mainly relies on manual inspection,and there are many hidden dangers such as missed detection,false detection,and slow efficiency.In recent years,the rapid development of deep learning has brought technological innovations to various industries,greatly improving work efficiency,and the security inspection field is also in urgent need of an intelligent security inspection device that can accurately and quickly identify contraband combined with deep learning.To address this problem,this paper proposes a lightweight contraband detection algorithm based on improved YOLOv5,which can be deployed in embedded devices with limited memory and computing resources for real-time detection.The main research work and results are as follows:（1）The imaging principle of X-ray and the workflow of security inspection equipment are studied and analyzed,to have a more comprehensive understanding of the task of contraband detection.Due to the particularity of X-rays,the generated images of contraband have problems such as serious overlap,various sizes,and missing features during the detection process.The collected X-ray images are processed through a variety of data enhancement technologies,and the number of images among various types of prohibited objects is balanced to obtain a high-quality X-ray security inspection prohibited object data set.A series of optimization measures have been carried out for the YOLOv5 algorithm given the difficulties of X-ray image detection and the requirement of a lightweight network.（2）In terms of lightweight improvement,the improved ECA is inserted into the Ghost bottleneck to form a GE module,which brings obvious performance improvement without increasing the complexity of the model.The network parameter specifications are obtained through experimental exploration,and a new lightweight backbone network is built,which improves feature reuse and detection performance.Other parts of YOLOv5 contain a large number of redundant operations,which are replaced by Ghost bottleneck and depthwise separable convolution,which significantly reduces the number of network parameters and FLOPs,by 43.8% and 26.7%,respectively.And the model size is only 7.97 MB.（3）In terms of performance improvement,the current mainstream activation functions are compared and analyzed,and the activation functions with excellent performance are added to the appropriate positions of various modules.For the missed detection and false detection in the detection process,the improved CA considering the channel and spatial information is utilized into the Neck to extract key and effective feature information.In addition,to improve the occlusion overlap in detection and obtain a more reasonable and effective prediction frame,optimization measures such as transfer learning,K-means++,and DIOU NMS were adopted in the model training stage,which further improved the network performance.The experimental results show that the constructed YOLOv5s-R performs better than YOLOv5 s,the m AP@0.5:0.95 achieves84.1%,and the detection speed also has 46.2 FPS,which can provide real-time detection.（4）The YOLOv5s-R is deployed in Jetson Nano,which is an artificial intelligence embedded device.Through the acceleration of Tensor RT,the detection results have reached a balance between accuracy and speed.