Research On Methods Of Apple Grading Under Different Color Light Source Based On Computer Vision

The development of apple grading method based on computer vision technologyhas considerable significance so far as advancing competitive power of Chinese applein the international market. At present, most apple grading systems have some defectssuch as lower accuracy rate, etc.According to some literature, the selection ofappropriate light source color helps to improve the detection accuracy and reliability ofagricultural product grading system. Therefore, apple grading with different light colorbased on computer vision was studied in this paper. The main contents are as follows:1. An all-direction apple images Acquisition experiment device with LED mixedlight source, which have simple structure and convenient operation, was developed.During the image acquisition process, the function of linear detection equipment wassimulated by apple sample rotating around its vertical center axis. Moreover the appleimage acquisition in a variety of color light source was preferably realized.2. In the HSI space, on the basis of contrastable analysis midpoint filteringmethod was selected to denoise, OSTU method to segment images. In addition,overlap area is reduced with Image cut and synthesis.3. Four kinds of features of apple images under different color light wereextracted, i.e. image shape, texture, red colored area and scars types. The grade modelsof single and multi-features were studied, respectively.4. Based on comparing different feature selection methods, a new featureselection method which combined Wilkes statistics with principal component analysisis presented. The effectiveness of the method was verified by grading results.5. The effect of different light color on feature selection method was compared.Among these methods, Wilkes statistic combining with principal component analysisin the yellow light had the best effect, the characteristic parameter numbers wasdecreased from32to11, FDA grading accuracy rate reached to98%, and BP neural network grading accuracy and prediction error were98.3%and0.0023respectively.The research results showed that the color of source light had significantinfluence on the feature extraction and selection as well as pattern recognition. Thechoosing of appropriate color light source and features could significantly improve theclassification accuracy and efficiency.