Recognition Of Antacid Bacilli Based On Object Detection

Antacid bacilli are the pathogens that cause tuberculosis and seriously damage human health.The detection can be done quickly using smear microscopy,but it requires doctors to observe and count under a microscope,which is a great workload.The antacid bacilli microscopic vision platform developed by the group can quickly acquire high-quality microscopic images and use built-in image processing technology to identify antacid bacilli microscopic images,which can reduce the burden of medical personnel and improve work efficiency.Traditional image recognition algorithms need to design features manually first and then use machine learning algorithms for classification,but such algorithms have low accuracy and poor robustness and cannot meet the recognition requirements for complex scenes.The object detection research based on deep learning lacks corresponding data sets,and the target objects of antacid bacilli are small in size and easy to overlap,and the background is complex and variable,so the object detection algorithm applicable to general objects is less effective for antacid bacilli recognition.Based on the microscopic vision platform,we established a dataset TB-data for antacid bacilli recognition,and then performed exploratory data analysis on the TB-data dataset to analyze the dataset in detail from the data perspective and provide data support for the subsequent data enhancement and algorithm design.A variety of classical general object detection algorithms are analyzed from the principle perspective,and the Cascade R-CNN algorithm is selected as the base algorithm,on which the antacid bacilli adaptation improvement is made.The improvements mainly include:a multi-sample fusion data enhancement strategy based on beta distribution is proposed,and a large number of new maps are generated using random scale overlay method mixup to make the model generalization ability stronger and avoid overfitting phenomenon.The sliding window cut strategy is proposed to cut the original high-resolution map into multiple sub-maps,change the relative size of the target and the original map,improve the recognition accuracy,and use the NMS algorithm to screen and merge and de-weight the duplicate regions after the recognition is completed;analyze some problems of the FPN feature extraction pyramid,and propose the MFMS（Multi-scale Feature Map Stack）method Repeatedly fusing high-resolution information and performing multi-scale information stacking to improve the detection of small objects of antacid bacilli.Multiple sets of detection algorithm comparison experiments are conducted based on the antacid bacilli dataset TB-data,including multiple general object detection algorithm comparison experiments,data enhancement comparison experiments,cut map strategy comparison experiments,FPN feature pyramid and MFMS module comparison experiments.The detection effects of different algorithms and the improved algorithms in this paper are compared under a variety of practical scenarios.The AP₅₀index of the improved model can reach 85.1%and the rate can reach 12.9fps,which is 6.4%higher than the AP₅₀of Faster R-CNN,with real-time,accurate and lightweight features,providing a new solution for the recognition of antacid bacilli microscopic images.