Comparative performance evaluation of fast discrete curvelet transform and colour texture moments as texture features for fruit skin damage detection

摘要

The paper discusses two approaches for fruit skin damage detection. In the former approach, two dimensional (2D) Fast discrete curvelet transform based texture features are computed. This approach divides image at fine level and curvelet transform is applied on each sub-image. The energies of these curvelet coefficients extracted from sub images are used as the feature vector. The later approach introduces a low level feature, colour texture moments which combines colour moments and local Fourier transform as a texture representation of fruit. In this approach, Local Fourier Transform is applied on image to derive eight characteristics maps for describing co-occurrence relation of pixel in various colour space and the first and second moments of these maps resulting in 48 dimensional feature vectors are calculated. Effectiveness of both feature vectors using classifiers namely Artifical Neural Network and Support Vector Machines are tested to sort defective fruits.