Research On Fruit And Vegetable Recognition Based On Computer Vision

China is the biggest producing country of fruit and vegetables. However, the total export of fruit and vegetables is far behind developed countries, lacking competency in the International market. In the bid for higher competitiveness and adaptation to people's daily life, the commercialization of fruit and vegetable is developing quickly in recent years, and the computer vision is becoming more and more widely used in fruit and vegetable recognition, classification and quality testing. Yet the large number of species and variation of shape, color and texture bring about the urge of lower feature dimensionality as well as a uniform method for various kinds of fruit and vegetables, which is one of the most important topics by now. For this reason, this paper presents a research on universal recognition method, which can apply on 35 varieties vegetable and fruit automatic classification. Its basic pricinple is as follow:firstly applies Gabor wavelet filtering on fruit and vegetable images to extract Gabor feature, and then performs dimensionality reduction by PCA and FLD, at last uses error correction SVM as the classifier. Experiments over 35 kinds of fruit vegetables are conducted, and proved the validity of the method, up to 97.14%. Research and improvements are carried on in the following aspects:This paper takes multi-class of fruit and vegetables instead of mono-class. Images of up to 35 kinds of common fruit and vegetable are collected, with 20 images of each kind, and finally a database of 700 images is established.For feature Extraction, since the geometric features of fruit and vegetables diverse immensely with low recognition rate, this paper adopts PCA, FLD and Gabor wavelet filtering instead. Firstly, features of PCA, FLD and Gabor are compared and experimentally verified. Then, the Gabor feature extraction of fruit and vegetables is studied. Kernel function of 2D Gabor can describe the receptive field of simple root cortex in the human visual system. Gabor feature of the fruit and vegetable image is fairly representative for the image. Since the Gabor feature of the fruit and vegetable image is very high, down-sampling is conducted before PCA and FLD to reduce the computation load.This paper adopts error correction SVM instead of traditional methods. Error correction SVM is a multi-class SVM classifier based on channel error correction coding. This paper adopts BCH (31,6) to correct up to 7 errors, i.e. if no more than 7 of 31 trained SVM got the wrong answer, they can be corrected. Therefore the recognition rate is further improved.