Study On Exterior Quality And Grading For Xinjiang Fuji Apple Using Hyperspectral Imaging Technology

Since recent methods for detecting and grading Xinjiang Fuji apple are backward, efficiency is low, and it is affected by subject ingredient easily. So the classification can’t be highly efficient and accurate base the same size and made the apples worse regularity. The study investigated the prediction performance of the exterior quality (including weight, size, and coloration area) of the apple using hyperspectral imaging technology, and on grading the exterior quality. Rapid detection methods and grading models were established on the basis of the study, which provides research and grading basis for apple quality online analysis using multispectral imaging technology.The hyperspectral image of the apple was acquired using hyperspectral imaging. The main contents and conclusions of the study on exterior and grading were list as follows:(1) The images of the characteristic wavelength were extracted. The image was used the image of single feature wavelength and the band ratio image base the two images of feature wavelength. The area features of object were extracted after segmenting image and the volume feature was gotten using the area feature. And then the predicted models were established. The best model was using two volume features and then was examined to be used predicting the protected sets. The correlation coefficient between predicted and real weight was0.993using the image of733nm wavelength feature in the first batch experiment, RMSEP=4.34g; the correlation coefficient using the852/713band ratio image in the second batch experiment was0.982, RMSEP=4.15g.It is obvious that the hyperspectral image can be usedfor predicting Xinjiang Fuji apple weight correctly and the image of single feature wavelength predict the weight correctly.(2) Firstly852/713band ratio image was segmented using threshold and removed stem’s region of binary image by morphological opening operation. And then the8-Connected Boundary tracking method was used to get the edge sequences. The Minimum Enclosing Rectangle method was subsequently used to get the size of the apple. Finally the regression model was established using predicted and actual value of the apple size. The results display that the hyperspectral imaging technology with the band ratio operation and the minimum enclosing rectangle method can be used to predict the size of Red Fuji apple, the correlation coefficient between predicted and real size was0.94, the maximum absolute error between the predicted and actual value of apple size was3.06mm, and root mean square error was1.21mm.(3) Firstly852/713band ratio image was segmented using threshold and removed stem’s region of binary image by morphological opening operation and that the complete apple segmented image was gotten. The three feature wavelength images was abstract based three basic wavelengths that were R(700nm), G (546.lnm), B(435.8nm) respectively, the hue image was gotten through the HIS translation. And then the accumulating of the gray value as the hue feature was abstracted from the hue gray image of the boundary defined by the binary image. The coloration area of apple was sorted using decision tree and AdaBoost_NN, respectively. The consistent rate of using decision tree and AdaBoost_NN to artificial classification attains99.2%and97.7%respectively.It is obvious that RGB image in the hyperspectral image can be usedfor sorting the apple coloration area.(4) Apple sample was sorted by manual method on exterior quality based the national standard. The classification feature was the apple to the size and the coloration area that were gotten by the before study. The apple was sorted using the Probabilistic neural network (PNN). The result shows that the consistent rate of using PNN to artificial classification attains82%.