Study On The Medlar Classification Method Based On Computer Vision

Medlar is a kind of traditional medicinal herbs in China, has many health benefits, In recent years, medlar is enjoyed by people as tonic, the medlar products are applied widely in the pharmaceutical industry and food industry and so on, and it’s demand has increased dramatically, so the quality requirements of the detection and monitoring for the medlar are increasingly high, therefore, there is urgent need to improve and enhance the detection method of the medlar origin and quality. At present, the testing method of medlar origin and quality mainly includes sensory identification, microscopic identification, chemical composition analysis and so on, but there are some limitation of the above discriminant analysis methods in a certain extent. medlar is difficult to enter the international market because there are lack of systematic and scientific identification method for the medlar quality. Therefore, in order to ensure the accurately distinguish the quality of medlar, establish a rapid, nondestructive detection system is a important research subject for the test of the medlar from different origin.With the development of computer technology, computer vision technology based on image processing has been widely applied in the detection of agricultural products quality because of its characteristics of rapid, nondestructive, convenient online detection etc, but the application of this method in medlar has not been reported. Therefore, in this paper, the medlars from different origin and with different quality grades are clasified; Through the research of visible light image and hyperspectral image, to establish the discriminant model, to provide the theory and practice reference for medlar grading. The mainly works as follows:The classification research of visible light image1. Collect the images of the 6 kinds of medlar samples(3 origin and 2 different grades for every origin) under 7 different colors of lights(including: red, blue, green, purple, yellow, cyan and white) using the image acquisition system developed in the laboratory, and extract the image feature parameters of color and texture.2. Through the theory of image information entropy to study on the effect of the different color light source on image quality, to choose the most beneficial image for medlar grading, result shows that, the white light is the optimum for the medlar image, the medlar classification accuracy up to 91.4% under the white light.3. In order to further improve the classification accuracy of medlar images under white light, The kernel Fisher disicrimination analysis(KFDA) classification method based on Wilks Λ criterion was put forward, the classification accuracy rate and validation reached 100% and 87.8% respectively.The classification research of hyperspectral image technology1. Using hyperspectral image acquisition system, collect six grade medlar samples hyperspectral images in the range of 371.05 ~ 1023.82 nm band.2. Through the theory of image information entropy to determine the characteristic wavelength is 950 nm, and extract the corresponding hyperspectral image to anslyze.3. Through extracted the texture feature parameters of the hyperspectral image, using the Fisher disicrimination analysis(FDA) discriminant analysis proved the effectiveness of the characteristic wavelength selected method based on the image information entropy. The result shows that the medlar classification accuracy up to 99.4% under 950 nm wavelength based on hyperspectral image, prove the effectiveness of the proposed method.4. Set up the FDA classification model based on spectral feature, and the classification accuracy rate and validation reached 100%.Compare the classification result of common images technology and hyperspectral technology, The FDA classification model based on hyperspectral image spectrum feature classification effect is the best. The result can provide theoretical reference for medlar grading based on computer vision.