Research And Application Of Small Object Detection Methods For Mycobacterium Tuberculosis

Tuberculosis affects human health and is highly contagious.At present,the screening of Mycobacterium tuberculosis is mainly based on the detection of sputum smears by light microscopy.The test standard is the statistical number of Mycobacterium tuberculosis in 300 visual fields of a single sputum smear,and manual operation is time-consuming and labor-intensive.With the continuous improvement of the deep learning model and the improvement of the detection accuracy,it is possible to apply the detection algorithm based on the neural network model to the detection of Mycobacterium tuberculosis.However,the background of the microscopic images of Mycobacterium tuberculosis is complex and the object is small.Most detection methods have problems such as weak model generalization ability and low recognition rate.This thesis starts from three parts:image enhancement method,object detection framework,and WEB system,and researches object detection of Mycobacterium tuberculosis.The main work is as follows:1.Create a Mycobacterium tuberculosis dataset containing 1700 microscopic images,use small object data augmentation methods on the training set,conduct a 1-3 times comparison test on the copy-paste method,and finally determine the 1 times paste method.This method improves Precision by 0.8%and Recall by 0.5%on the original basis.2.Propose an Anchor-Based small object detection framework based on a bidirectional feature fusion pyramid.Based on the Faster R-CNN network,the ResNeXt structure is first introduced.Then the feature fusion pyramid is introduced,and a new structure BiCrFPN（Bi-directional Cross-scale Feature Pyramid Network）is proposed.In addition,the DIOU bounding box regression loss function based on the center point distance is adopted.The above method finally reaches 89.2%Precision and 94.7%Recall.3.Considering the balance of detection accuracy,detection speed,and model size,propose an attention-module-based AC-CenterNet small object detection framework.The Hourglass-104 structure is first optimized,and the original residual module is replaced with the BottleNeck structure.At the same time,a depthwise separable convolution is used to replace the standard convolution of the residual module.This operation reduces the number of parameters by 28.5%based on only reducing the detection accuracy and speed by a small amount,making the model lightweight.Finally,the ECA module with a small number of parameters is introduced,which improves Precision and Recall by 0.6%and 0.3%respectively without affecting the detection speed and parameter amount.4.Design a tuberculosis detection system based on B/S architecture,which includes a front-end web page module,a server-side module based on the Flask framework,and a database module.This system has good availability and is helpful for the promotion and popularization of tuberculosis detection.