FPGA implementation of the principal component analysis algorithm for dimensionality reduction of hyperspectral images

Implementation of the Principal Component Analysis onto High-Performance Computer Facilities for Hyperspectral Dimensionality Reduction: Results and Comparisons

“…Specifically, in [10], Fernandez et al present an implementation of PCA Jacobi on a Xilinx Virtex XC7VX690T FPGA, which is evaluated with two of the images used in this research work: AVIRIS Cuprite (large) and AVIRIS Jasper Ridge. The main characteristics of the FPGA targeted in the aforementioned work are summarized in Table 11.…”

“…Since the criterion of the number of processing cores does not make sense in the FPGA case, only the operating frequency is included in the comparison metric. This comparison is presented in Table 12 and Figure 10, and it has been conducted using two AVIRIS images -Cuprite (large) and Jasper Ridge-as input, since they are the ones analyzed in [10]. conducted using two AVIRIS images -Cuprite (large) and Jasper Ridge-as input, since they are the ones analyzed in [10].…”

“…This comparison is presented in Table 12 and Figure 10, and it has been conducted using two AVIRIS images -Cuprite (large) and Jasper Ridge-as input, since they are the ones analyzed in [10]. conducted using two AVIRIS images -Cuprite (large) and Jasper Ridge-as input, since they are the ones analyzed in [10]. Although the most efficient implementation is the FPGA one, as its operating frequency is smaller when compared with the other two (76 MHz), it is worth noting that this advantage is achieved at the expense of a smaller number of CPS when compared to a GPU and a design-time and power consumption increase (more complex programmability and up to 60 W of typical power consumption [31]) when compared to an MPPA.…”

“…Equation (9) summarizes the complete Jacobi method, where K represents the total amount of iterations, C K is a diagonal matrix containing the eigenvalues of the input image, and P i are the successive Jacobi rotation matrices. Similarly, the associated eigenvectors can easily be computed, as shown in (10), where the eigenvectors are stored in a column-wise order.…”

“…Additionally, when updating the columns (Code 5.3.3), matrix P is also updated following Equation (10). The methodology to compute these products is very similar to the one that is described in (8), as each element to be zeroed only modifies four elements of its associated P matrix.…”