A 28.8-mW Accelerator IC for Dark Channel Prior-Based Blind Image Deblurring

This work presents an accelerator that performs blind deblurring based on the dark channel prior. The alternating minimization algorithm is leveraged for latent image and blur kernel estimation. A 2-D Laplace equation solver is embedded to reduce the latency by 56% for boundary wrapping. For latent image estimation, gradient data locality is employed to reduce the latency by 57%. A sorting engine is designed to reduce the latency in data access by 96% for calculating the dark channel. A pipelined mixed-radix 1-D fast Fourier transform (FFT) engine is used for efficient latent image estimation and blur kernel estimation. By employing image size approximation, 85% of additions and 97% of multiplications for FFT can further be saved. In the blur kernel estimator, a 2-D convolution engine with a parallel architecture is implemented, reducing the latency by 79%. The accelerator supports blur kernels of
$25\times 25$
and
$49\times 49$
pixels for blurred images of
$129\times 129$
and
$257\times 257$
pixels, respectively. Fabricated in 40-nm CMOS, the accelerator’s core area is 3.98 mm2. The chip dissipates 28.8 mW at 65 MHz from a 0.65-V supply. It can estimate a blur kernel of
$25\times 25$
pixels for an image patch with
$129\times 129$
pixels for deblurring a full-HD image in 1.7 s, achieving a 2562
$\times $
shorter latency than a high-end CPU. Compared with the state-of-the-art design, the chip achieves a four times higher normalized area efficiency and a 7.5
$\times $
higher normalized energy efficiency.