TEXTURE CLASSIFICATION BY IMPLEMENTING BLUR, SCALE AND GREY SHIFT INSENSITIVE TEXTURE DESCRIPTOR BASED ON LOCAL FOURIER TRANSFORM

초록

In this paper a new descriptor based on local Fourier transform called the local circular difference phase patterns (LCDPP) is proposed for medical image analysis. LCDPP is designed from the blur invariant property in the Fourier domain and its performance is tested with blurred, rescaled and grey value shifted images. LCDPP utilizes the phase information obtained from the differences of circular neighborhood pixels and the center pixel. Four complex coefficients of the resulting Fourier transform are transformed into 8-bit binary code. A histogram of the resulting code words is created and used as the feature for the texture classification. Experimental results show that the average accuracy in classifying original texture images of LCDPP is about 4% better than that of the well-known LBP method while the average classification accuracies in classifying blurred, rescaled and grey value shifted texture images for LCDPP are about 50%, 60% and 10% better than those of LBP, respectively.