Study On External Defects Classification Of Navel Orange Based On Machine Vision

Machine vision is widely used in quality detection and automatic grading of agricultural products today. Applying machine vision technique and decision tree technique of multi- characters, different kinds of external defects of navel oranges under static condition were studied in this research. Color, shape and texture characters of different defects were acquired first, and then a decision tree based on multi-characters was constructed to classify external defects of navel oranges. Then the images obtained under dynamic condition were researched. Main contents and results were as follows:1. The research advancements and achievements in the area of nondestructive inspection of fruit quality and automatic grading, especially the research on fruit external defects detection, using machine vision technique were reviewed and the existing problems were put forward.2. Machine vision system for this research has already been used in real producing line. The system is composed of a grabber, a CCD camera, 6 fluorescent lamps and an ADVANTECH INDUSTRIAL COMPUTER 610 computer.3. R/B, R/G were used to segment the surface defects of fruits. When R/B was set as 2.0 and R/G was set as 1.5, it could lead to a good result. R/B value of defects was below 2.0 and R/G value of defects was below 1.5, while the values of good ones were both above the two thresholds. It is also showed that using this method to segment defects will get a satisfying result.4. The wrong pixels information which occurred when segmenting the background and defects was recovered in this research, using the pixel information around the wrong pixels in the image. It is showed that two algorithms worked very well.5. Color, shape and texture characters of defects were extracted. The color character included R, G, B, H, I, S, R/G, R/B, (R-G)/(R+G), (G-B)/(G+B), (B-R)/(B+R), (2R-G-B)/4 and so on. The shape character was composed of length, width, length/width, area, perimeter, compactness, the total number of defects, the percent of defects and so on. The Texture character included entropy, energy and so on.6. The characters of total, percent, R/G, LAV were used to constructed decision tree. A method was used to decide which character was used first to construct the decision tree. After the decision tree was constructed successfully, it was used to grade five different kinds of normal defects on navel oranges surface, including crack, rot, stem, sick spot and ulcer. The result showed that their singular accuracy were 85%, 80.5%, 100%, 91.9% and 100%, and their total accuracy were 89.5%, 85.3%, 80%, 91.9%, 98.3%.7. Images were obtained under dynamic condition. The detecting accuracy of good navel oranges and the oranges with defects based on color character were 92% and 88%, respectively. The reasons for the faults occurred were also analyzed in this study. In a word, the results were acceptable.