A Lightweight Method For Multi-target Detection In Cardiopulmonary Imaging Based On Dynamic Bagging Integration Strategy

The main target objects of biomedical image processing are medical images in various image formats.Currently,the type of biomedical images used clinically is mainly X-ray images(X-CT).In the current imaging medical examination,most of the lesions can only be seen through two-dimensional CT images,which usually can only be diagnosed through a doctor’s examination.As China gradually enters into an aging society,it is obvious that doctors have a heavy workload,which may affect the accuracy of detection.Therefore,the design of a computer-aided medical imaging detection system has broad prospects for application in disease detection,and it has guiding significance for doctors to diagnose diseases.At present,the existing medical image detection methods can only identify a single type of disease,but cannot detect multiple types of diseases,and the existing computer vision algorithms are not lightweight,and the detection process is also a test of hardware resources.In recent years,deep learning models have performed well in terms of accuracy and lightweight,and convenient new algorithms have also provided new ways and perspectives for medical image detection.Aiming at the shortcomings of current medical image detection,this paper proposes a deep learning method based on Bagging ensemble learning for medical image detection.The main work is as follows:In terms of the ensemble learning framework,a dynamic weighted Bagging ensemble learning method(DW-Bagging)to solve the imbalance of the number of various samples is proposed.In view of the imbalance in the number of various diseases in the data set in this paper,the original Bagging ensemble learning medium-weight voting method expands the model variance and generalization error.This paper proposes a confidence-weighted voting Bagging integrated learning strategy to replace equal-weight voting.The optimized integrated learning method DW-Bagging reduces model variance,improves model generalization performance,and improves model detection accuracy.In terms of improving the detection performance of the base learner,it is aimed at the problems of low efficiency,time-consuming,long receptive field,small receptive field and low precision in the detection process.Firstly,YOLOv5 is selected as the base learner for integrated learning,but the original YOLOv5 backbone feature extraction network adopts the C3 structure,which makes the model parameters redundant,reasoning and analysis efficiency is low,and it is difficult to apply to some highly integrated industrial production scenarios.In order to realize the lightweight of the network model and strike a balance between speed and accuracy,we adopted a method of replacing the backbone network and introducing the lighter Efficient Net Lite network into the model.This replacement method makes the model have fewer parameters,improves the detection speed,and achieves a more efficient target detection task without reducing the detection accuracy;in order to improve the receptive field of the model,this paper proposes a content The feature pyramid network configuration(CARAFE-FPN)that completes feature recombination by means of perception solves the problem of missing small target information due to inconsistent target scales in the feature fusion process;finally,the design of the loss function,in order to improve detection accuracy and accelerate convergence,use boundary The box loss function SIo U Loss,adding the Angle penalty cost effectively reduces the total degree of freedom of the loss,which greatly helps the training convergence process and effect.This paper conducts experiments on the chest image CT data set,converts the data format,preprocesses and constructs a new data set.In this paper,multiple sets of bounding box loss functions are set up for comparative experiments,and the SIo U loss function with the best performance is selected.Improvements are made from two aspects of the weak learner and the integrated learning framework.m AP and F1 score are selected as the accuracy evaluation indicators of the base learner,FPS and Params are used as the lightweight evaluation indicators of the weak learner,and Accracy is used as the overall performance of the integrated learning model.Compared with the existing detection methods,this paper adds lightweight feature extraction and upsampling modules on the basis of dynamic integration boxes.The experimental results show that the m AP of the method in this paper reaches 44.1% in terms of weak learner accuracy,the F1 score reaches 45.8%,the FPS in terms of lightweight is 64.3f s,the Params is 14.4MB,and the accuracy rate reaches 81.69% after the dynamic weighted model is integrated.Therefore,the method in this paper is It has more advantages in the detection of multi-target chest medical images.