I am using wavelet as feature descriptor. When I use small size subband like 16x16 then getting better result than using 32x32 sizes subband. When I take large size then I am losing edge information. Why?
as you are using small size wavelet i.e. small scale, it is use to detect high frequency in your signal. if you use wider wavelet, it will not detect the high frequencies in 1D or 2D signal.
Your question is a little bit fuzzy. The reason is that you are using maybe the wrong type of wavelets for your objects. It has nothing to do with your window size itself. Please make sure that you use the coorect wavelets for your edge detection. If your wavelet length is too long you will definitely not detect an edge in either direction. Your then have to take a look in other scales. Here is a paper which describe the effect of different wavelets on scales. In the references you will find more papers.