Apple Stem/Calyx And Defect Identification By Their Texture Feature

Fruit is one of the main sources for human to obtain nutrition which are important to health. The requirement of fruit quality is improved because people have a higher requirement on the quality of life. Consumers estimate fruit quality mainly by their external quality, therefore, detection of fruit external quality becomes necessary for fruit industry of our contury, and it is also a significant techonology for the commercialization treatment ability for post-harvest fruit and improving their competitiveness in the world market. There has been a problem in fruit external quality detection, that is how to indetifiy the stem/calyx from the defect, especially those with similar size and shape. Hyperspectral technology, multi-spectral technology, multi-camera system and3-D image analysis method were applied to solve this problem. Some of the drawbacks of those methods include large volume of data, low efficiency and high cost. This research was aim to find a simple method for stem/calyx and defect identification, improving correct results while combined with other mehods, and could be used for on-line process in future.The research objects were apples. Texture of apple stem/calyx and defect were analysed using statistical and wavelet methods. Texture feature vectors were obtained by caculating the mean values and the standard variance values of the characteristic parameters of GLCM or the high-frenquency wavelet coefficiences. Support vector machine was used as the classifier.The main research contents and results were:(1) Texture features of apple stem/calyx images and defect iamges were analysed. The obtained results indicate that:DT-CWT method can obtain more comprehensive texture information than GLCM method; the influences of image rotation are both small to these two mehods;3-level DT-CWT is the optimal deformation wavelet in the research.(2) Three image classification methods (K-means, BP nueral network and LS-SVM) were compared. The classification results indicate that for texture image classification, LS-SVM is better than K-means and BP nueral network, and the RBF kernel and one-against-one algorithm are the the optimal kernel function and multi-classification algorithm for texture images classification. Apple stem/calyx and defect texture images were analysed by3-level DT-CWT, and the correct classification rate was92.22%by the LS-SVM classifier.(3) GLCM method and DT-CWT method were used analyzing stem/calyx and defect images, the correct classification rates were81.17%and92.22%respectively, and the computation time were0.25s and2.22s per image for these two methods. The results indicate that DT-CWT is a better texture analysis method than GLCM method when applied for identification of apple stem/calyx and defect image.