Research On Classification And Intelligent Defects Detection Of Endoscopic Images In City Sewer

The city sewers are widely distributed,with many kinds,complex disease types and different evaluation criteria.It is difficult to determine and repair subsequent diseases.The security and universality of endoscopic image pipeline detection have been greatly improved,and digital image processing technology has successfully turned the automatic identification and monitoring of several pipeline diseases into reality.But pipeline endoscopy image processing methods generally have some problems,such as small application scope,low recognition accuracy and heavy workload,so it is urgent to develop a new automatic recognition method.This method divides the problem into two parts:(1).Classification of endoscopic images in sewers;(2).Image detection of diseased parts.For image classification,endoscopic images are divided into 18 categories,including irrelevant images,conventional images and various kinds of disease images,according to the real detection video.On this basis,considering the processing method of mixed images,Inception-ResNet-v2 is transformed into a convolutional neural network which can realize multi-label classification,and the complete classification of endoscopic images in pipelines is realized.Considering not only identifying the location of the disease body,but also accurately segmenting the disease area,this paper uses Mask RCNN model for reference,combines the good performance of YOLO-v3 network in target detection,and puts forward Mask-YOLO-v3 model for different types of diseases.These models are not responsible for classification,but only detect the target in the classified pictures and provide the smallest disease area.At the same time,the mask of the disease area is detected,which makes the detection result more intuitive and clear.In the image classification experiment,comparing the classification performance of MLL-Inception-ResNet-v2 with ResNeXt-101 and SEENet-101,the results show that the classification mAP of the model is 70.1%,and the relative classification accuracy of the model is 0.9% and 0.4% higher than that of the control model,respectively.In thedisease detection experiment,comparing the detection performance of Mask-YOLO-v3 and Mask-RCNN,the detection mAP of the former is 81.4%,compared with that of the former.The comparison model is improved by 0.4%.Based on the above research results,a comprehensive detection platform is implemented to detect 17 common classified pipeline diseases according to the existing standards.The results show that the platform can achieve more than 97% detection accuracy in linear working time compared with the manual observation method,which widens the application scope and greatly improves the detection efficiency.