Research On Cocoon Image Recognition Based On Machine Vision

The variety of cocoons greatly determines the silkiness of silk,and ultimately affects the economic benefits of raw silk in the market.At present,the cocoon sorting mainly relies on the identification of the worker to complete the manual sorting,but the manual sorting takes time and effort,and the quality of the identification depends heavily on the subjective consciousness of the selected workers.In order to stably and objectively identify and classify cocoons and reduce labor intensity,it is imperative to study the image recognition of cocoons based on machine vision.However,the existing cocoon image processing research still stays in the identification of single cocoon.It is difficult to apply to the visual part of the cocoon separating robot.In order to overcome these problems,this paper studies the cocoon image recognition based on machine vision.Because of the existing cocoon contour extraction algorithm is difficult to extract the spotted contour accurately.It is proposed to use the DRLSE(Distance rule level set evolution)model based on least squares method to extract the cocoon contour.Then,according to the extracted contour information of cocoons,the cocoons are positioned to realize the identification of multiple cocoons.Constructing a deep learning model to complete the identification of cocoons to avoid the loss of image information caused by artificial extraction of cocoons.The paper mainly completes the following work:(1)The contour of cocoon as one of the important features of cocoon identification,the extraction precision will greatly affect the correct rate of subsequent recognition.In order to facilitate the subsequent contour extraction,an image enhancement algorithm based on the double-filtering anti-sharp mask is proposed,which improves the contrast of the cocoon edge and make the cocoon edge relatively smoother.(2)Because of this paper identifies multiple cocoons and use a low resolution and black-white camera,using the Canny operator is difficult to accurate extract the spotted cocoon contour of the spot at the edge.So introducing the DRLSE model commonly used in medical image segmentation into the field of cocoon image segmentation.DRLSE model based on the least squares method is proposed,aiming at the over-fitting and under-fitting problems of the DRLSE model in the evolution of the cocoon contour.The recognition accuracy of GA-SVM trained by improved DRLSE model is 93.8479% on test set.Compared with GA-SVM trained by DRLSE model,the recognition accuracy of GA-SVM trained by improved DRLSE model on test set is improved by 2.3334%.The validity of the algorithm is verified.(3)The feature of manually selecting features is cumbersome and may neglect the characteristics of the object that can be reflected.This paper uses two deep learning models,GoogLeNet and se-googlenet,and compares their recognition rates.The correct recognition rate of SE-GoogLeNet on the test set is 98.27%.The recognition accuracy rates on the same test set was 4.4221% higher than GA-SVM and 0.65% higher than GoogLeNet.