Research On Automatic Recognition Theory And Technology Of Formed Components In Leucorrhoea Microscopic Images

Leucorrhea routine is one of the routine means of female physiological examination.The main routine examination method of leucorrhea is through microscopic imaging of leucorrhea secretions,and the corresponding index parameters are obtained by identifying and counting the visible components under the microscope,so as to analyze the pathological characteristics of the samples to be tested.Because leucorrhea routine has non-invasive characteristics,patients are easy to accept.In China,tens of thousands of women undergo routine leucorrhea examinations every day.However,at present,the routine detection of leucorrhea is performed manually by doctors,which is easy to cause cross-contamination,and the detection efficiency is very low.In addition,samples with inflammation generally have a strong odor,which seriously affects the working environment of medical staff and easily leads to drug resistance.With the development of machine learning and deep learning technology,leucorrhea routine develops to automation and intelligence,and the most critical is the automatic detection and recognition algorithm of the tangible components of microscopic images.There is no related research on the recognition of multi-type components in leucorrhea microscopic images at home and abroad.However,the recognition of cells or bacteria in microscopic images is limited to the recognition in specific environments.Aiming at the characteristics of the visible components in leucorrhea microscopic images,the related theories and key technologies involved in the detection of the visible components are deeply studied.The main achievements and research contents are as follows:1.Detection of epithelial cells in leucorrhea microscopic images based on texture featuresEpithelial cells in leucorrhea microscopic images are an important component of leucorrhea microscopic images.In view of the large area and reticular texture of epithelial cells in leucorrhea microscopic images,a digital image processing method for forground location is proposed,which is used to extract the foreground targets such as epithelial cells and impurities in microscopic images.According to the extracted foreground target,the texture features(local binary pattern(LBP),Gabor and histogram of oriented gridients(HOG))of foreground target are analyzed and compared,and the texture features of target are classified by support vector machine.Finally,the best recognition result is obtained by using LBPfeatures and HOG.Experiments show that the proposed method has fast detection speed(304 ms),high detection precision(86.9%),and low miss-detection(recall: 88.9%),which meets the real-time detection requirements of the routine leucorrhea.2.Small target detection and recognition based on improved R-CNN in leucorrhea microscopic imagesFor the identification of other visible components in leucorrhea,such as white blood cells,red blood cells and fungi,a target detection method based on improved R-CNN is proposed in this dessertation.Firstly,a region proposal method for leucorrhoea microscopic images is proposed,which is 50 times more efficient than the selective search in R-CNN model,who produces fewer foreground targets,which greatly reduces the complexity of object foreground extraction in R-CNN.Secondly,a foreground target classificaiton and box regression method based on Inception-ResNet-V2 network is proposed.The collected data are used to verify that the method achieves high precision and recall detection of the above-mentioned tangible components,and the mean average precison is 79.75%.3.Detection of adherent cells in leucorrhea microscopic images based on improved Faster R-CNNFor leucorrhea samples with high cleanliness,there are many visible components such as cells in their views,and the components adhere to each other.Using the methods above can not achieve the segmentation of these adhesive visible components.Aiming at these problems,a target detection method based on improved Faster R-CNN is proposed.Firstly,based on the analysis of morphological features of the tangible components,an improved region proposal network is proposed with less anchors,which means more effective in training and testing.Secondly,a dimension reduction method based on PCA is proposed to normalize the dimension reduction and dimension of the feature map,and input it into the fully connection layer for classification and box regression.Through the test of experimental data,the method can effectively segment and recognize adherent cells,and the algorithm has fast execution efficiency and high accuracy.4.Trichomonas detection method in leucorrhea microscopic images based on VIBE improved modelAccording to the fact that trichomonas is still active in fresh leucorrhea samples,a trichomonas detection method based on VIBE improved model is proposed from the point of view of moving target detection as it is difficult to detect trichomonas based on the morphological methods.The algorithm improves the way of updating the background model in VIBE,and the spreading mechanism of neighborhood background model.At the same time,a series of impurity filtering measures are proposed.The experimental results show that the algorithm can effectively improve the accuracy of trichomonas detection and other indicators.Compared with other moving target detection algorithms,the best detection results are achieved.