Fast Gabor texture feature extraction with separable filters using GPU
Wai-Man Pang,Kup-Sze Choi,Jing Qin
DOI: https://doi.org/10.1007/s11554-013-0373-y
2013-09-17
Abstract:Gabor wavelet transform is one of the most effective texture feature extraction techniques and has resulted in many successful practical applications. However, real-time applications cannot benefit from this technique because of the high computational cost arising from the large number of small-sized convolutions which require over 10 min to process an image of 256 × 256 pixels on a dual core CPU. As the computation in Gabor filtering is parallelizable, it is possible and beneficial to accelerate the feature extraction process using GPU. Conventionally, this can be achieved simply by accelerating the 2D convolution directly, or by expediting the CPU-efficient FFT-based 2D convolution. Indeed, the latter approach, when implemented with small-sized Gabor filters, cannot fully exploit the parallel computation power of GPU due to the architecture of graphics hardware. This paper proposes a novel approach tailored for GPU acceleration of the texture feature extraction algorithm by using separable 1D Gabor filters to approximate the non-separable Gabor filter kernels. Experimental results show that the approach improves the timing performance significantly with minimal error introduced. The method is specifically designed and optimized for computing unified device architecture and is able to achieve a speed of 16 fps on modest graphics hardware for an image of 2562 pixels and a filter kernel of 322 pixels. It is potentially applicable for real-time applications in areas such as motion tracking and medical image analysis.
computer science, artificial intelligence,engineering, electrical & electronic,imaging science & photographic technology