Image classification is an important research field of computer vision

and the key problem of which is to select a type of feature and establish the similarity metrics between images. In view of the mural image characteristics

the contour feature plays an important role in expressing the mural image semantics. Many studies have shown that the contours can be used as an important feature in image recognition and classification. However

previous studies tend to use the chamfer distance between each pair of the most similar contours to compute the similarity between images

or build local descriptors for each contour

clustering into codebook

and describe the image features as histograms

then do the classification using SVM. However

these methods ignore the overall structure between the contours

lack of the overall view of the all contours

while in reality the semantics of an image tend to be more of a holistic semantic. In this paper

we study the similarity metrics between images based on the overall structure of contours

the calculation of contour similarity is subject to the constraint of the space structure relations with other contours

the generated similarity are more able to express the overall similarity between two images. The experimental results show that our method improved accuracy compared to others in mural image classification.