The Research On Automatic Identification Of Butterfly Species Based On The Digital Image

Automatic identification technology of insect is very important for solving the prombles (such as short of hands, inefficiency, high cost, Poor universality and so on) in the traditional classification methods. In recent years, with the rapid development of computer vision technologys and pattern recognition, developing the automatic identification system is a new research field on identification of insect species. Butterfly is an important branch of the insect world; it has a great variety of species (more than 14000) and a wide distribution, so the identification of it was a very hard and complex work. Moreover, the larvas of butterfuly are the major pest to the agriculture and forestry. They will have a terrible effect on the foods and environment.In this paper,750 butterfly specimen images (15 images of each) form 50 butterfly were used for automatic identification. On the one hand, we researched the morphological characteristics and the textural features of butterfly; on the other hand, we proposed a cascade classifier for automatic identification of butterfly.For getting efficient features extraction, we proposed a new method:Histograms of Multi-Scale Curvature (HoMSC) to describe the contour of butterfly; and for representing the textures of butterfly, a new method:Gray-level Co-occurrence Matrix of Image Blocks (GLCMoIB) was implemented. HoMSC can be used for reflecting the concave-convex changes of the contour and it is more effective and accurate than other methords (which used in some correlational researches). GLCMoIB is base on the traditional Gray-level Co-occurrence Matrix, but the difference is our methord not only reflect the global changes of the image, but also can be used for describing the local variations of the texture (which are the important distinction in different butterfly species).For implement an automatic identification system of butterfly, an improved Weight-Based k-Nearest Neighbor search algorithm was used for classifier design. Under normal conditions, the K.NN algorithm has a good performance in processing classification problems of high dimensions and small samples, and its results depend on the species which has the most number in feature space. However, according to the analysis of our testing results, we found that the distance in feature space was a very important element for the final decision. Therefore, we tried to use a weigh vector to control the effect of the neighboring points according to the distance from the center point. In addition to improve the final identification rate and keep the the speed of the system, we designed a cascade classifier.With these methods, a butterfly classification system was implemented on Windows, which developed with the visual C++ and Qt4. It worked on 50 buttery species and the final identification accuracy is 98%. The experimental results show that the proposed method works well for identication of buttery species with images of arbitrary orientation and scale.