Reserach On Micro-vision Detection Method For Surface Cleaning Quality Of Panax Notoginseng

As one of the largest biological resources in Yunnan,the cleaning quality of Panax notoginseng is the basic premise and guarantee for the follow-up medical deep processing of Panax notoginseng products.According to the on-the-spot investigation of Panax notoginseng production enterprises,at present,the cleaning process of Panax notoginseng mainly relies on manual visual and experience to judge its cleaning quality.However,due to the heavy laterite texture and the complex shape and surface texture of Panax notoginseng,it is difficult to effectively guarantee the cleaning process quality of Panax notoginseng only by visual inspection.In view of the limitations of the macro-inspection of the cleaning quality of Panax notoginseng,this paper proposes a micro-vision method for fine inspection of the cleaning quality of Panax notoginseng surface.The key technology of image segmentation in micro-vision inspection is studied to improve the existing manual visual inspection technology and ensure and improve the deep processing quality of Panax notoginseng products.The main work of this paper is as follows:1.To develop an image acquisition scheme for on-line sampling and micro-detection of Panax notoginseng cleaning,collect micro-images of Panax notoginseng surface under IMAGINGSOURCE DFK41BU02 industrial microscope system,establish image data set,totally 150 frames of images,and complete manual labeling at pixel level.2.Based on matconvnet deep learning framework,a full convolution neural network for image target region segmentation is constructed.image enhancement technology is used to enhance the edge information of the image,so that the target region has better separability.The task of segmentation of residual target area of soil on the surface of Panax notoginseng was preliminarily realized.3.In view of the complex texture features and obvious color features of Panax notoginseng surface,a multi-color feature fusion segmentation algorithm based on FCN8 s network model structure is adopted.The method of transforming the input image into the color space is used to fuse the color features of the output image,which achieves the segmentation task of the residual soil target area on the surface of Panax notoginseng and improves the accuracy.4.Aiming at the characteristics of too few training samples of Panax notoginseng surface micro-image,the ablation model of convolution neural network was built independently.Based on the Ablation Experiment of independent color space,the microvision detection model of Panax notoginseng surface cleaning quality was established for different color space.5.The segmentation algorithm based on the self-built best independent color space ablation model is implemented to predict 40 test images in the surface micro image data set of Panax notoginseng.The experimental results show that the best ablation model of independent color space in this paper achieves good experimental results.The average pixel accuracy is 82.27%.In this paper,an Panax notoginseng cleaning device is designed as an image acquisition scheme,and the micro-visual image data set of Sanqi surface is established.The residual soil region segmentation of Sanqi surface based on full convolution neural network and multi-color space fusion is successfully realized.It provides a method basis for extending the visual inspection algorithm to the production line of Panax notoginseng cleaning process to guarantee the quality of deep processing of Panax notoginseng products.