Rapid Detection Of Mycobacterium Tuberculosis Based On Lightweight YOLOv4

Tuberculosis is a chronic respiratory illness caused by mycobacterium tuberculosis infection,which can cause serious impact on human body.The prevention and control situation is very serious.Due to the complex characteristics of tuberculosis lesions,some traditional medical diagnostic methods are expensive,have an impact on human body,false detection and omission.At present,the combination of deep learning and medical technology has gradually become an important research direction of medical diagnostic assistive technology,but there are still some problems,such as lack of public data set and inconsistent data set format.With the deepening of the network,the structure of many detection models is relatively complex and the number of parameters is huge.It is not applicable to small and medium-sized medical institutions,community censuses,etc.,where there is a lack of highly configurable equipment,and is difficult to deploy to mobile devices such as embedded devices.In order to address the above issues,this paper proposes an improved lightweight YOLOv4 pulmonary tuberculosis detection model,and the specific work contents are as follows:(1)Expansion of data set.To achieve better training results,a data augmentation method is used on the input side of the model,and then the dataset is then expanded and annotated with random geometric transformations and optical transformations to allow the model to learn the features of the target more adequately during training,enabling multi-scale training.(2)Lightweight design of YOLOv4.Replacement of CSPDarknet53 size-consistent feature layers in the YOLOv4 backbone feature extraction network using the linear bottleneck structure of MobileNetv2.A significant reduction in the amount of network parameters and computational effort is achieved by changing the underlying network architecture and replacing the standard convolution with a deeply separable convolution.(3)Improved lightweight YOLOv4.The lightweighting of YOLOv4 dramatically reduces the computational effort and number of parameters of the model,while leading to a decrease in detection accuracy.Improvements to the model’s enhanced feature extraction network with some effective training strategies are considered.Firstly,average pooling is introduced in the SPP layer to reduce the problem of target loss caused by maximum pooling,and more target information can be retained in scale fusion.Secondly,the fusion with shallow feature layers is increased in PANet from three to four scales,so as to solve the multiple scale convergence problem in target detection,enhances the ability of the model to extract small target features in the data set,and improve the detection accuracy.Experiments show that the lightweight design of the YOLOv4 and the improvement of the enhanced feature extraction network,compared with the YOLOv4 model and other target detection networks,only sacrifice a small number of precision,the model can make a significant reduction in the number of parameters and the amount of computation,not only better detection of small targets,and at the same time the FPS is also improved,which can realize the rapid detection and identification of Mycobacterium tuberculosis.