The Research On Enhancement Algorithm For Low-illumination Image

Enhancement processing need be implemented to improve the quality and visual effect of the low-illumination images. There are two traditional methods to enhance the low-illumination digital image:the methods based on frequency and the method based on time. The histogram enhancement, the most popular way to enhance, is representing the latter one;while the methods based on frequency often uses wavelet transform, FT(flourier transform), DCT(discrete cosine transform)to enhance. This research mainly discusses the immune genetic algorithm and wavelet neural network enhancement of the methods based on frequency.Image signal will be decomposed enough by using the good local characteristics of wavelet transform.The energy of the image redistributes after wavelet transform.Most energy of the image locates in the part of low frequency (wavelet coefficients is larger).The wavelet coefficients will be enhanced after wavelet transformation. The enhanced image will be get after reverse transformation.The algorithm of image enhancement based on wavelet neural network is introduced. The basic model is based on the BP neural network model of wavelet image enhancement. The main idea is adjusting the value to minimum the error. Gradient search technology is adopted in BP neural network, so as to minimum the actual output value and mean square error. In the learning process, the error transmits back and corrects the value at the same time.The immune genetic algorithm (IGA) based on the Euclidean distance and elite cross is introduced based on the BP neural network model.The image enhancement method based on IGA is proposed. The mechanism of various maintaining and immune memory is introduced to improve the global search ability. Experiment proves:the IGA enhancement algorithm is better than the traditional BP neural network enhancement algorithm in stability and convergence rate.There are two issues in this processing.One issue is the selection of the wavelet coefficient. Biorthogonal wavelet function with the characteristic of linear phase is selected to get higher reconstruction accuracy and prevent the image distortion after enhancement. Consider the standard deviation of the original image:larger standard deviation leads to higher frequency. The longer filter is needed,so the support length of the filter must be wider. The support length of the filter for decomposition and for reconstruction is 2Nd+1 and 2Nr+1,so Nd and Nr must be larger. Experimental results verifies above analysis.Another issue is the selection of the times of the wavelet coefficient. To achieve this target, we introduce the neural network, and use the BP neural network to select and then get the best one. The neural network has a character of self-study, Therefore the enhance system has a self adapting ability. IGA introduced in improves the study rate.