Research On Grading Of Apple's Shape Based On Machine Vision

Quality grading of fruit is an important item in post-harvest handling and marketing, it is mainly performed according to its shape, color, size and various kinds of defects and blemishes. Fruit grading has been always performing manually in China, which has drawbacks such as inobjectivity, inconsistency and inefficiency.Technical advancements in the areas of Computer Image Processing and Machine Vision have helped the movement towards automatic grading of fruits. This study mainly developed a research work on the algorithm about the extraction of the apple's stem, extraction of the feature of the apple's shape, design of artificial neural network classifier based on BP & GA algorithm and segmentation of the bruising part. The algorithms are programmed in Visual C++ language.Background condition for fruit grading is simple. On the basis of stretch transform of gray level, gray image can be binarized ideally if given a overall gray segmentation threshold. As stem is very small compared with apple, the shape of apple can be segmented if morphology is used, synchronously the stem can be extracted.Firstly, Fourier Descriptor used to describe the object's shape is introduced to the extraction of feature of apple's shape. The boundary of apple is an obturated curve, distance between boundary point and centroid r(k)' is used as a function to express the 2-D boundary in spatial domain. The Fourier transform coefficients of the discrete sequences F(h) are spectrum values in frequency domain, and they compose feature parameters of shape.The r(k)' signature can be reconstructed by Invert Fourier Transform. The No. 1, 3, 4, 5 harmonic in the spectrum based on radius represent almost all shape information of apples. A 3 layer-neural network is used as classifier, the number of nodes in hide layer is determined on trial and error.Based on gradient descent rule, the BP (Back Propagation) algorithm is a local optimization algorithm. Genetic algorithm has the sufficient ability to optimize in the whole scope, in this dissertation they are combined as learning algorithm of neural network. 72 samples of 3 classes are used to train the neural network, 42 test samples proved that accuracy over 73 percent can be achieved through this approach and it is invariant to translation, orientation and scale variance.Judging apple from its single image is inadequate, in this dissertation two image which are acquired from two directions are used, that is to say, the latter image is acquired after the apple is rotated an angle of 90.Bruising on the surface is a common defect of apple, in the saturation image of HIS color model, gray value of defect varies from the normal part, defective part can be segmented out when 90 is choosed as threshold.