Orange Outer Quality Detection Based On Multi-Scale Geometry Analisys And Machine Learning

Orange is one of the principal fruits in china which is with high product, easy to grow and has good economic returns; but in China, the orange which have been treated after post-harvest only accounts for 1% of gross production. In comparison, the production of orange been post-harvest tested in developed countries accounts for 90% of total production. In comparison with developed countries, China's agricultural economy is very weak and the income of farmers is very low. So the majority of farmers can not offer the expensive testing equipments.As a result, export oranges from China are only a small number of the total orange exports in world and the export price is less than the half of the world average price. Therefore, the study of orange quality detection method with the characteristic of fast testing speed, high accuracy and easy operation is very urgent. In this paper, the orange outer quality detection is carried by the orange infrared spectroscopy and the orange image character based on multi-scale geometry analysis and machine learning. The main content of this paper includes the following fields:(1)Orange image pre-processing. For the purpose of acquiring the high accuracy in orange image segmentation and edge detection, it is necessary to denoise the original orange image first of all.Combining the noise feature of orange image, the second generation curvelet transform is used to decompose the original orange image and deal with the high frequency coefficient by the square neighbour threshold method. The experiment result shows that the denoising method brought forward in this paper can efficiently eliminate the noise in the original orange image and it's capability is better than the wavelet transform de-noising algorithm and the adaptive threshold de-noising algorithm.(2) Orange image segmentation. In this paper, differential evolution algorithm and Ostu algorithm is used to segment the de-noised orange image. In order to compensate for the brink of fracture or the loss of critical information by separately using segmentation algorithm, the improved canny operator is utilized to edge detection of original orange image. Then calculating the segmented and edged image, by or operator to obtain the final segmented image.(3)Orange outer quality detection. After orange image edge detecting and segmenting, the size,shape and surface damage extent of orange can be acquired according to orange image character. Such as some orange image with not obvious damage surface, we can via the fourier transform infrared spectroscopy of orange outer surface,support vector machine and continuous wavelet transform to identify the orange outer quality.80 couples of FTIR are used to train and test the proposed method, where 60 couples of data are used as training samples and 20 couples of data are used as testing samples. The feature vector is input to support vector machine (SVM) to train so as to accurately classify the good orange and rotten orange.As the experiment result, the proposed method has a high recognition rate to the good oranges and rotten oranges.