| The research of signal and image processing has far-fetching theoretical significance and extensive application value. As one of the branch of data compression technology which belongs to the subject of computer science and technology, image data compression technology has extensive application fields and rapid development rate. But there are many difficult points in this subject. Wavelet analysis developed quickly from the mid of1980s. Different from Fourier transform which only contains frequency domain information, wavelet transform contains both frequency and time domain information. The time-frequency analysis characteristic of wavelet analysis can help it play a significant role in the field of signal and image compression, which includes:wavelet vector quantization compression, the wavelet spectrum analysis of EEG signals etc. From the1980s, artificial neural network (ANN) which involves many different advanced disciplines has been arousing wide concern of all kinds of researchers such as mathematicians, physiologists etc. ANN has an extensive application in real life which contains signal and image processing, system identification etc. Owing to excellent characteristics such as good approximation abilities, self-learning and self-adaptive abilities, ANN can be used signal compression.At present, the combination of wavelet analysis and ANN is an active research area. There exist two kinds of combination types:loose and compact. The former one uses wavelet analysis as ANNs’pretreatment, while the latter one uses wavelet basis functions as ANNs’activation functions and this structure can be called wavelet neural network (WNN). In addition, considering the fact that the link between the outputs of WNN is linear, it avoids the problem of local minimum which is widespread in traditional ANNs. This paper presents a method of image compression based on WNN. It also gives priority to the following three cruxes:the initialization of parameters, the adjustment of parameters and the choosing of learning rate. Image compression by WNN and many different kinds of experimental are conducted in this paper, and the comparison between WNN and traditional ANN is presented here. The result shows that WNN succeeded in improving performances and efficiency in image compression. It also shows that the methods of initialization, adjustment of parameters and the method of choosing learning rate discussed in this paper can guarantee the stability of the WNN. |
