Block Region of Interest Method for Real-Time Implementation of Large and Scalable Image Reconstruction

Abstract

In field programmable gate array (FPGA)-based reconstruction of high-resolution (several-gigabyte) images, architectures that store the entire source image in block memory are no longer feasible. In this letter, a block region of interest (ROI) method is proposed for large and scalable image reconstruction, using 2D $4 \times 4$ truncated sinc interpolation by storing only a slice of the source image in block memory. To ensure a high hit rate in block memory for interpolation and improve memory access efficiency, this method provides an effective criterion for partitioning, which subdivides the output image into blocks based on the ROI in the source image. Moreover, a special storage pattern is designed to simultaneously provide 16 pixels for each interpolation to enable real-time implementation by full pipelining with no memory data redundancy. The proposed method was validated by experimental results on an FPGA board along with the corresponding simulations on MATLAB. Various image sizes were tested to prove its flexibility. The time required to reconstruct an output image of $32~\hbox{k} \times 32~\hbox{k}$ at 200 MHz was 2.79 s, thereby achieving a processing speed of 385 Mpixels/s in single floating point precision.