Image super-resolution aims to obtain a high-quality image at a resolution that is higher than that of the original coarse one. This paper presents a new neural network-based method for image super-resolution. In this technique, the super-resolution is considered as an inverse problem. An observation model that closely follows the physical image acquisition process is established to solve the problem. Based on this model, a cost function is created and minimized by a Hopfield neural network to produce high-resolution images from the corresponding low-resolution ones. Not like some other single frame super-resolution techniques, this technique takes into consideration point spread function blurring as well as additive noise and therefore generates high-resolution images with more preserved or restored image details. Experimental results demonstrate that the high-resolution images obtained by this technique have a very high quality in terms of PSNR and visually look more pleasant.
We present a competitive learning vector quantization with evolution strategies for image compression. This technique embeds evolution strategies (ES) into the standard competitive learning vector quantization algorithm (CLVQ). After each iteration during the CLVQ training process, the so-far generated codebook is adjusted by the embedded ES through its recombination, mutation, and selection process. The proposed algorithm can efficiently overcome CLVQ's problems of under-utilization of neurons and initial codebook dependency. The embedding of ES greatly increases the algorithm's capability to avoid local minimums, leading to a global optimization. Experimental results demonstrate that it can obtain significant improvement over CLVQ and other comparable algorithms in image compression applications, especially when it involves larger codebooks.
This paper presents a learning based codebook design algorithm for vector quantization of digital images using evolution strategies (ES). This technique embeds evolution strategies into the standard competitive learning vector quantization algorithm (CLVQ) and efficiently overcomes its problems of under-utilization of neurons and initial codebook dependency. The embedding of ES greatly increases the algorithm’s capability of avoiding the local minimums, leading to global optimization. Experimental results demonstrate that it can obtain significant improvement over CLVQ and other comparable algorithms in image compression applications. In comparison with the FSLVQ and KSOM algorithms, this new technique is computationally more efficient and requires less training time.
A neural network based image enhancement method is introduced to improve the image resolution from a sequence of low resolution image frames. Most of the existing methods reconstruct a high-resolution image from a multiple of low-resolution image frames by minimizing some established cost function using a mathematical technique. This method, however, uses an integrated recurrent neural network (IRNN) that is particularly designed to be capable of learning an optimal mapping from a multiple of low-resolution image frames to a high-resolution image through training. The IRNN consists of four feed-forward sub-networks working collectively with the ability of having a feedback of information from its output to input. As such, it is capable of both learning and searching the optimal solution in the solution space leading to high resolution images. Simulation results demonstrate that the proposed IRNN has good potential in solving image resolution enhancement problem, as it can adapt itself to the various conditions of the reconstruction problem by learning.