| With the development of deep learning in computer vision,medical image target detection has become one of the hot topic.Tuberculosis(TB)has been included among the top ten leading causes of death worldwide,but the traditional method of manual screening for tuberculosis bacilli has vast workload and high misdetection rate,ofen relying on clinician doctors’ s subjective judgement.Due to a series of problems in sputum smear image such as complex background and small target of TB bacili,TB bacili target detection is a difficult challenge.To solve the above problem,a small target algorithm for TB bacili based on improved SSD(Single Shot multibox Detector)network is proposed in this thesis.By inputting the original sputum smear image obtained by autofocus scan into the improved SSD network,the object recognition and location of TB baculli were completed.The main works of this thesis are as follows:Based on the analysis of the difficulties in the target detection of TB bacilli,SSD network was selected as the target detection algorithm of TB bacilli,and the SSD network structure and loss function were improved respectively.Aiming at the problem that SSD network has poor detection effect on small targets,an improved SSD algorithm based on SPP-Net(Spatial Pyramid Pooling Net)is proposed,named SE-SSD.And the SSD structure improved the performance on small target detection of TB bacilli by combining the low-level feature and the high-level feature.The SE-SSD network achieves 78.7%m AP(Mean Average Precision)by training and testing the data set of TB bacilli target detection.Compared with SSD network,it is improved by 2.3%.Aiming at the problem that the number of imbalanced of training data over positive and negative samples,the location is not accurate and the classification effect is not ideal.In this thesis,Focal loss function is introduced to improve SSD loss function,named SEF-SSD.So that the networks pays more attention to training positive samples and hard samples.The SEF-SSD network achieves 79.8%m AP by training and testing the data set of TB bacilli target detection.Compared with SSD network,it is improved by 3.4%.In this thesis,SSD algorithm is selected as the target detection algorithm of tuberculosis bacilli.An improved SSD algorithm named SEF-SSD based on feature fusion and Focal loss was proposed in this paper to improve the detection effect of small targets of tuberculosis bacilli.The experimental results show that the improved algorithm has certain reference value for the target detection of tuberculosis bacillus. |
