A TENSORFLOW BASED METHOD FOR LOCAL DERIVATIVE PATTERN

Abstract

Python programming language provides a very convenient environment of implementing machine learning applications. However, programmers usually faced with a poor performance compared to compiled functions when they write script based programs that demands intense computations. TensorFlow framework provides acceleration by enabling the utilization of various computing resources such as multicore CPU and GPU unit as well as including various compiled algorithms for developing machine learning applications. In this way, algorithms developed using existing TensorFlow operations can shorten computation times by using these resources indirectly without requiring parallel programming or GPU programming. In this study, Local Derivative Pattern (LDP) analysis which is one of the efficient feature extraction approaches for machine learning models was realized using a TensorFlow based algorithm. Independent pixel based operations in LDP algorithm which requires intense computations, enable developing an efficient tensor based algorithm. The performance of the TensorFlow based algorithm has been measured by comparing it with the Python script version of the same algorithm. The results obtained for various sizes and numbers of sample images show that TensorFlow operations provide significant acceleration rates for the LDP algorithm.