decomposition and reconstruction of signal
discrete wavelet transform
scale (approximation) and detail wavelet spectrum
level of wavelet transform
smoothness of wavelet basis (number of vanishing moments for wavelet)
high-frequency (high-pass) and low-frequency (low-pass) filter
point and extensive image details
The problem of extraction and enhancement of image details of different sizes has been investigated. As experimental data aerophotos (video stream) received from an avionic IR optoelectronic system are used. Image-processing is used for solving 2 general (typical) issues:
1) extraction and enhancement of point objects of frame approximate to 1x1 pixel dimensions;
2) extraction and removal of extensive, low-frequency image details, which contain a background nonuniformity of the frames.
Image-processing is implemented by filtering of the image frequency spectrum (content). Discrete Wavelet Transform (DWT) is used for image orthogonal transformation.
The time-frequency representation is a more appropriate (useful) method for solving of the assigned tasks. A group of the time-frequency representations provides spectral details information localized on signal time base or its coordinate axis. DWT is the most promising sort of time-frequency representation in the current technology.
Issue 1: extraction of point objects of an IR frame is implemented by multi-step reconstructing of detail DWT coefficients only. Approximation DWT coefficients are truncated in the reconstruction. The procedure repeats until the reconstruction matrix will reach the source frame size. The bandwidth of this HF image filter is tuned with a number of DWT levels used in processing (computing).
The wavelets with the highest smoothness and FIR filter lengths of the function are the most efficient in solving issue 1. For example: Coiflets Wavelet and Discrete Meyer Wavelet.
Issue 2: extraction of background nonuniformity of IR frame, is implemented by multi-step reconstructing of approximation DWT coefficients only from a specified decomposition level. The reconstruction also repeats until the computed result will reach the source frame size. In this way a LF image-filtering is released, which conveys low-frequency frame details into a separate image. Then the obtained LF pattern is subtracted from the source image. The processed image is more useful for following contrast enhancement.
Wavelet-processing of presented plots shows good results and provides simple hardware-software implementation. However, the proposed method has a set of disadvantages: as evident decrease of dynamic range of the processed image; and appearance of low-frequency, low-intensive artifacts on it, by reason of spectrum aliasing effect.