Image improvement and data reduction techniques using neural networks

This research is aimed at developing efficient image processing techniques using neural networks. In this dissertation, a number of new neural network based techniques are presented for image enhancement, image restoration, superresolution and image vector quantization. Images are often impaired by noise. To deal with a mixed Gaussian and impulse noise in the image, a neural network based nonlinear filter was developed. This filter is composed of a region smoother, an edge detector and a synthesizer. By using specially designed neural networks, the filter is capable of efficiently removing the additive noise while preserving the image details.; Images can also be degraded (blurred) due to the imperfection of the imaging system. In this dissertation, a new image restoration technique was developed to provide an improved restoration process and also an enhanced image quality. This technique reformulates the restoration problem to include an evaluation function that combines a scaled residual term with a space-variant regularization and achieves an optimal solution of the problem through a Hopfield type neural network.; Another problem in image processing applications is that images are often undersampled and have a low resolution due to the use of an imperfect image acquisition system. This dissertation presents a novel neural network based method for image resolution enhancement. This method utilizes a particularly designed recurrent neural network capable of learning and searching optimal solutions in the solution space for optimization problems. Simulation results demonstrate the good potential of this method in solving image resolution enhancement problem with a fast computational speed.; Images, especially digital videos, usually require a large, even huge, storage and transmission bandwidth. A vector quantization algorithm was designed for image compression to reduce this requirement. The algorithm embeds evolution strategies into the codebook generation process and can efficiently handle CLVQ's problems of initial codebook dependency and neuron under-utilization. Simulation results show that this algorithm can obtain significant improvement on overall performance over some other comparable algorithms.; While image compression can reduce the requirement of storage and transmission bandwidth of image, images often suffer from coding artifacts when they are encoded at a low bit rate. This dissertation also presents a new postprocessor for image coding artifact removal. The capability of this postprocessor is achieved through an adaptive FIR filter in which the filter coefficients are automatically generated by neural networks tuned to local image features through learning. Experimental results demonstrate its efficiency in reducing the coding artifacts.