Research On Quality Detection Method Of Infiltration Drip Irrigation Pipe Belt Based On Machine Vision And Deep Learning

Drip irrigation and infiltration irrigation are currently advanced water-saving irrigation methods.The patch drip irrigation belt and infiltration irrigation pipe are two efficient water-saving irrigation products.The quality of the punched holes in the production line directly determines the quality of the product.The quality problems of the hole position include missed punching,multiple punching,poor location,punching and poor shape,etc.In addition,the body of the infiltration irrigation pipe is made of a mixture of various particles.If there is a problem with the injection molding process,it will cause surface defects such as weld marks.At present,the production line quality detection method is inefficient and has a high rate of missed detection,making it difficult to meet the requirements of high-quality production.This project relies on the Yangtze River Water-saving Equipment Technology Research Institute of Jinan University,combined with the needs of enterprises,and uses machine vision and deep learning technologies to complete the research on the quality detection method of infiltration drip irrigation pipe belt.Including the quality detection method of the patch type drip irrigation belt hole position and the surface defect detection method of the infiltration irrigation pipe.The research content of this paper is as follows:(1)Improvement on the current detection method of patch drip irrigation belt hole position and the Open CV and depth edge detection method Dexi Ned technology are used to complete the preprocessing,edge detection and postprocessing of the drip irrigation pipe image,and then use the Hough transform to complete the positioning of the key position,and finally judge the hole whether it is qualified or not,compared with the current hole position quality detection method,this method has better edge extraction effect and faster detection speed.The disadvantage is that when using the deep learning edge detection method,the construction cost will be higher;(2)Applying deep learning object detection to locate key positions in drip irrigation belts.Established a patch type the drip irrigation belt dataset is collected with a Nikon camera,a total of 662 drip irrigation belt image data,including two types of images of the patch drip irrigation belt body and key positions.Applying YOLOv1,YOLOv3 and YOLOv5 object detection algorithms,in which YOLOv1 uses modules such as Res Net network and Bottle Neck CSP;YOLOv3 and YOLOv5 respectively use Dark Net53 and CSPDark Net53 as the backbone network,and all training uses data augmentation methods to improve the model detection effect,including multi-scale,mixup etc;in terms of recall,the recall of YOLOv1 and YOLOv3 is not ideal,and the recall of YOLOv5 using auto-anchor reaches 99%;in terms of m AP,m AP of YOLOv1@[0.5,0.95] is 55,the metric m AP@[0.5,0.95] of YOLOv3 and v5 is 73.Since YOLOv5 s has a near-perfect performance in terms of Recall metric,it is finally determined that YOLOv5 s is a better method for locating the key positions of patch drip irrigation belts,and the detection speed of the model fully meets the real-time requirements.Finally,based on the key position positioning method,a method for judging the quality of hole position based on proportion is proposed;(3)A method for surface defect location and quality judgment of infiltration irrigation pipes based on YOLOv5s-ECA is proposed.Firstly,established a dataset of infiltration irrigation pipes,the dataset contains 1408 pictures,a total of five categories,through the introduction of ECA(Efficient Channel Attention)module,which reduces the complexity of the detection model and improves the training accuracy.The C3ECA＿X module is formed on the basis of the original C3＿X module of YOLOv5,and based on the C3ECA＿X structure and ECA module,the YOLOv5 s network structure is modified,including the improvement of the backbone network and feature fusion network,finally,a variety of network structures are finally formed.From the experimental results,most of the modified structures improve the m AP@[0.5,0.95] metric by about 1.5%.In addition,SWATS,which has both Adam training speed and SGD generalization ability,is used in the experiment.For the optimizer,it can be seen from the metric such as Precision,Recall,and m AP that the final detection model has a strong positioning ability.Finally,based on the location of the key position of the irrigation pipe,a kind of multi-type defect judgment method of infiltration irrigation pipe is proposed;(4)Based on Py Qt5,developed a quality detection system for infiltration and drip irrigation pipe belt,including database system,operation interface and related detection functions,etc.,and conducted black box and white box tests.