Research On Fish Species Classification Algorithm Based On Computer Vision Technology

There are many different kinds of freshwater fish in China, different kinds of freshwater fish has different nutritional value and market value, and many species of freshwater fish similarity by the naked eye can’t identify. At present in the domestic processing industry to mainly depends on artificial fish species classification, classification of artificial fish species affected by man-made factors, not only low efficiency but also affect the survival of the fish cycle, and need a lot of workers, stimulated the economic costs. With the development of science and technology, appeared the mechanical classification of fish equipment, the classification of mechanical damage to the fish is relatively large, and can only be grading the size of the fish, fish species can not be accurately identified and judgment. With the rapid development of computer visual technology, classification of fish by using computer vision instead of our eyes, which is undoubtedly a good way to solve the above problems, this thesis is to realize the automatic classification of fish species, for the purpose of establishing a nondestructive classification of fish species model based on computer vision technology.This paper selected six common freshwater fish as the research object, and build a freshwater fish collected images of experimental platform for computer vision.According to the feature extraction of fish, this paper discussed the image preprocessing algorithm suitable for fish, In order to better segmentation fish and background, weighted average method is adopted to improve the fish of the gray level;For the purpose of the binarization is to highlight the characteristics of the fish, the commonly used threshold processing method, based on noise in the image, an improved adaptive median filter to remove method;In order to extract the geometry characteristics of the fish, the boundary tracking method was adopted for contour extraction.Image preprocessing is for the purpose of feature extraction, this paper in order to fully reflect the characteristic information of the fish and extracted the three major characteristics such as color, texture, shape, a total of 24 characteristic value;Compare the color of six kinds of intuitive description, can not distinguish between each fish very well, by extracting RGB and HSV color components mean and variance of the data, help through the data to distinguish each fish; In view of the different kinds of fish, fish of texture have very big difference, here with the method of gray level co-occurrence matrix, the extract of texture characteristic value of five; Fish special geometric shape is irregular and uneven density,which need for comprehensive extracting the fish of the area, perimeter, long axis and short axis features to reflect the geometrical characteristics of the fish.This paper collected samples of 120 fish, and characteristic parameters of 24 of these samples the principal component analysis method is used for dimension reduction,extract the four principal component can reflect all the information in the fish, the aim is to use lesscharacteristic information reflecting the whole fish, to reduce the data amount of calculation.For small sample classification algorithm, this paper chose the classification algorithm based on support vector machine(SVM), in order to improve the recognition rate,selection cross-validation grid optimization method of parameter optimization, chose the radial basis kernel function, and in the interval [0, 1] normalization of data, the final classification recognition rate is 95%, classification of reached the expected goal.