Study On The Detection Methods Of Skin Pathological Specimens

In the traditional process of submitting skin pathology specimens for examination,due to the need to go through different stages of submission and the large number of specimens,the loss rate of specimens during the examination is relatively high.Therefore,the batch detection of skin pathology specimens can not only prevent the loss of specimens and facilitate rapid self-examination after loss,but also has great significance for the intelligent management of specimens and the construction of smart hospitals.Therefore,this thesis takes skin pathology specimens as the main research object,and studies the detection methods of skin pathology specimens.The main research work includes the following aspects.First,analyze the characteristics of the research object and actual application requirements in this thesis;then compare the detection networks such as Faster RCNN,YOLOv3,YOLOv5 and YOLOX-Dark Net53,and choose YOLOX-Dark Net53 as the basic network;Then optimize it on the basis of it,introduce the auxiliary network and attention mechanism to extract the regional information of skin pathology specimens,and solve the problem of fewer available features of the target to be detected due to acquisition occlusion.Finally,the fusion factor is used to control the information passed from the deep network to the shallow network,so that the output of each layer is more focused on the feature learning of this layer.The experimental results show that the m AP of the improved YOLOX-Dark Net53 model is 99.47%,and the FPS is 29.Compared with the original YOLOX-Dark Net53 model,the m AP is increased by 2.24%,and the FPS is basically equal,which shows that the model is effective in skin specimen detection tasks.In order to further improve the detection speed,a two-stage skin pathological specimen detection method is proposed,which improves the overall detection speed by performing different processing on detection targets of different sizes.The first stage is based on the YOLOv3 model to detect the specimen bottle storing the specimen.First,for a single-size target to be detected,the calculation amount of the model is reduced by optimizing the detection scale of the model;second,the regression loss of the model is optimized to improve the regression accuracy of the model;finally,the model divides the batch of multi-sample images into a certain number of one-shot images.The second stage uses the output results of the first stage to construct a dataset of skin pathology specimens,and uses the lightweight network Mobile Netv2 to classify single-sample images and output visualization results.The experimental results show that,compared with the improved YOLOX-Dark Net53 model,the model reduces the reasoning time for a single image by3 ms when the detection accuracy is equivalent.