In this paper we present a novel Adaptive Compressive Sensing (ACS) for natural images, named M-ACS (Modified Adaptive Compressive Sensing). Using our
algorithm an image is divided into blocks and a Discrete Cosine Transform (DCT)
coefficients of each block are then selectively calculated (sampled) directly from the
raw data. Our ACS algorithm uses a novel prediction mechanism to select which
coefficients should be sampled, based on a novel stochastic power-law model for the
distribution of DCT coefficients and a uniquely designed noise estimation method.
As opposed to standard Compressive Sensing (CS), the number of coefficients to be
sampled is not predetermined and is not identical for all blocks. The underlying assumption we use is a power-law model for the distribution of the DCT coefficients
in each block. Several coefficients are sampled and used to estimate the block distribution model, which is then used to determine if additional samples are needed
to achieve a desired image quality. Unlike many research papers in the field which
test their algorithms on processed images, we do all our tests using raw image data.
We demonstrate through several examples, the performance advantage of our ACS
over standard CS algorithms, both in reconstructed image quality, bandwidth and reconstruction computational simplicity. Using our algorithm a reduction in the power
consumption required by an imaging device can be achieved.