| Since entering the 21 st century,with the continuous progress of science and technology,a large amount of information generated in people’s work and life has been used effectively.The development of computer hardware and software has made Artificial Intelligence widely used in various industries,such as education,medical care,and industrial design.Biometric identification technology is commonly used in information security and other fields,relying on its advantages of security,uniqueness,and difficulty of forgery.Iris recognition is a typical research topic of biometric technology,it is often used in identification,entrance guard systems,and attendance checking systems,but iris recognition systems are generally used in embedded systems or mobile devices,which have limited storage space and computational performance,so the recognition system is required to be convenient and efficient.Lightweight network model is one of the effective ways to solve this problem.Iris image object detection,as the first step in the iris recognition process,requires accurate extraction of iris and pupil,so the performance of the system is directly affected by this work.Combined with the need for ease of deployment of the iris recognition system,a lightweight model for iris image object detection is proposed in this paper.The model in this paper is improved by using the YOLOv5 model as the basic framework,taking advantage of the computational strength of the inverted residual module in the Mobile Net V3 model,and using this module to build a lighter feature extraction backbone network,replacing the conventional convolutional layers and C3 modules in the original backbone;and embedding the SRM module after the deep convolutional layer of the inverted residual module,which can improve the problem of a large number of parameters in the SE module when the number of feature map channels increases,and the global standard deviation pooling is introduced to improve the problem of inadequate capture of global information in the feature map by the SE mechanism and enhance the feature extraction capability of the model.In the field of model training,the non-maximum suppression algorithm for prediction frames is studied.By combining the linear confidence reduction strategy in Soft-NMS and the overlap calculation method incorporating the width and height loss of the prediction frames,the prediction frames screening process is optimized to improve the possible object misses in the detection process.To summarize the above scheme,a lightweight iris image object detection model is proposed in this paper.In this paper,multiple sets of model detection experiments and comparison experiments are set up to perform a more adequate experimental analysis.In the model detection experiments,the lightweight improved model proposed in this paper is trained and tested on several datasets to evaluate the detection accuracy,detection speed,and object detection visualization of the model,proving that the model can achieve considerable results in the iris object detection task.The ablation experiment and algorithm comparison experiment are also set up.In the ablation experiment,each of the above optimization schemes is disassembled to obtain multiple fusion models,while the above SRM attention is embedded into the original YOLOv5 structure to enrich the experimental reference model to make the strategy of control variables more complete;in the algorithm comparison experiment,algorithms such as YOLOv4 are introduced.The models are trained and tested on multiple datasets,and the analysis of the main indicators shows that the models in this paper achieve the m AP of 0.9925 to 0.9945 on multiple datasets,the number of parameters reaches 28.25 M,which is 47.3%less than the original YOLOv5 structure,the average object detection time is reduced by 43%.Through experimental analysis,the model in this paper has a more balanced performance on all indicators,which proves that the optimization scheme proposed in this paper is effective. |
PDF Full Text Request