Distance transform network for shape analysis

Shape is known as an important source of information in object analyzes and has been studied for many years for this context. In the object classification task, several challenges such as variations in rotation and scale, noise, and degradation make the problem even harder. This paper proposes the Distance Transform Network (DTN), which combines the power of networks and the richness of information provided from Euclidean distance transform for shape analysis. First, a distance map is obtained by the application of the Euclidean distance transform on each contour. Thus, each radius of dilatation is modeled as a network. Then, degree measurements of the dynamic evolution network are used to characterize the contour. Finally, a robust feature vector is composed by characteristics of different radiuses of dilatation. The methodology was tested in seven benchmarks available databases, including two otolith and three sets containing shape of leaves species which presents challenging contours with a lot of intra-class variations. The results against literature methods show that the proposed DTN is effective for natural shapes classification according to the higher success rates obtained in all cases. The advantages of our approach include robustness to degradation and noise, and tolerance to variations in the shapes scale and orientation.