Dedicated Hardware for Real-Time Computation of Second-Order Statistical Features for High Resolution Images

Abstract

Abstract. We present a novel dedicated hardware system for the extraction of second-order statistical features from high-resolution images. The selected features are based on gray level co-occurrence matrix analysis and are angular second moment, correlation, inverse difference moment and entropy. The proposed system was evaluated using input images with resolutions that range from 512×512 to 2048×2048 pixels. Each image is divided into blocks of userdefined size and a feature vector is extracted for each block. The system is implemented on a Xilinx VirtexE-2000 FPGA and uses integer arithmetic, a sparse co-occurrence matrix representation and a fast logarithm approximation to improve efficiency. It allows the parallel calculation of sixteen co-occurrence matrices and four feature vectors on the same FPGA core. The experimental results illustrate the feasibility of real-time feature extraction for input images of dimensions up to 2048×2048 pixels, where a performance of 32 images per second is achieved.

Parallel implementation of Gray Level Co-occurrence Matrices and Haralick texture features on cell architecture

“…Many researchers have been working on accelerating the process of computation the GLCMs and texture features extraction algorithms on FPGAs platforms [5,11,[14][15][16][17]. Tahir [5] presented an FPGA-based coprocessor for GLCM and texture features and their application in prostate cancer classification.…”

“…Bariamis et al [15] presented a hardware implementation to calculate 16 cooccurrence matrices and 4 feature vectors using a single core. The implemented hardware exploited both the symmetry and the sparseness of the matrix They chose N g = 32, the number of gray level for different image sizes from 512 × 512 to 2048 × 2048.…”

“…They used 5-bit gray-level representation just like [15]. Bariamis et al [16] calculated both the GLCMs and features in hardware, while a significant part of the computations relies on software.…”

“…Bariamis et al [16] calculated both the GLCMs and features in hardware, while a significant part of the computations relies on software. Their design in [15] allows the computation of GLCM features in hardware, but employed data redundancy in order to achieve high processing throughput. However, the redundancy led to high memory capacity requirements and redundant transfers of data over the PCI bus.…”

“…First, some implementations such as [5] required large external memory banks, while some processing is performed by a host machine. Second, other implementations such as [11,15] included symmetry and sparseness matrices, which is not a general implementation to support all kinds of images. Finally, these implementations calculate GLCMs without implementation considerations for improving the performance of the Haralick texture features.…”