Local feature recognition for industrial radiation imaging with DWT and SVMs

A radiation image local feature recognition algorithm based on SVMs (support vector machines) was designed and developed. Using a set of 4000 simulated images, we achieved at least 93.4% detection rate and 0.8% false positive rate. The employment of wavelet, in particular DWT (discrete wavelet transform), introduces multi-resolution support to the algorithm and increases total recognition performance. DWT decomposes a radiation image into a total of 3d+1 subbands and stresses features of different scales in each subband. Because of the standout generalization capability of SVMs, using subbands directly as input becomes possible and produces compelling results. In this paper, different kernel functions are also compared to each other. Experiments show that Gaussian radial basis function (RBF) kernel overtakes the others in our application.