Efficient computations of DSP problems using artificial neural network techniques

Abstract

The Digital Signal processing ( DSP) is an important area of research which is widely applied to many useful applications like telecommunication, image processing, instrumentation, biomedical, geophysics etc. The problems related to these applications have been solved by conventional methods which are non adaptive and inefficient. Recently developed Artificial Neural Networks (ANN) have proved to be important and efficient tools for solving many complex problems. An artificial neuron is a computational processing element which contains inherent non-linearity. It is adaptive in nature and resembles to the human brain cells. ANNs are the collections of such neurons embedded in a distributed manner in the network. In the present thesis some burning problems o f DSP have been attempted to be solved by ANN technique. The first problem considered is a very simple problem o f computing the log and antilog o f decimal numbers. For log computation, a single layer - single neuron ANN structure has been suggested where as for antilog computation which is more complex being an inverse operation, a functional expansion ANN has been proposed. In both cases, the proposed structures predict the results which are almost the same as may be obtained from the log tables. The computation o f the Discrete Fourier Transform (DFT), Discrete Hartley Transform (DHT) and the Discrete Walsh Transform (DWT) and their respective inverses is a useful process which is applied in many DSP applications. In the next problem these transforms with their inverses have been computed using the simple LMS technique using a single layer network without involving any nonlinearity. Here the outputs predicted by the proposed LMS technique match exactly with the actual outputs. The convolution and the deconvolution are important tools in geophysics, communication and image processing. In the third problem undertaken in the thesis, designs o f a linear and a circular deconvolver have been proposed in the thesis.