Bilateral two-dimensional least mean square filter for infrared small target detection

Parallel software implementation of recursive multidimensional digital filters for point-target detection in cluttered infrared scenes

“…Other approaches to the problem of point-target detection in cluttered infrared scenes tend to either focus on 1-D temporal filters [15], or 2-D spatial filters [16], [17]. When 3-D approaches are considered, different logic is typically applied in the spatial and temporal dimensions [18].…”

Small target detection based on accumulated center-surround difference measure

“…Background estimation based small target detection method is widely studied in recent years [6][7][8][9]. These methods detect small targets in the residual image which subtracts the estimation image from original image.…”

Multidimensional digital filters for point-target detection in cluttered infrared scenes

“…blue sky); however, this is likely to produce a high probability of false alarm for dynamic backgrounds and a low probability of detection for static targets. On the other hand, background estimation and subtraction algorithms or other high-pass filtering frameworks, operating in two-dimensions (2-D) on each frame in isolationsuch as Wiener filters 2 , least-mean-squares filters 3,4,5 , top-hat transforms 6 , moving average filters 7 , median 7 and bilateral 3,7 filtersclearly do not suffer from these problems; however, the powerful discriminants of temporal coherence and disparity, which are essential cues in biological vision systems, are lost. Some methods attempt to solve this problem using one type of 1-D filter in the temporal dimension and a different type of 2-D filter in the spatial dimension 8 .…”

“…The 2-D moving-average prediction-error filter and the 1-D polynomial prediction-error filters could indeed be regarded as being limiting cases of the proposed approach. In the former case, only a direct-'current' (DC) spatial component with one-frame temporal support and wide-area spatial support is considered; whereas in the latter case a higher- 3 order model is used with one-pixel spatial support and temporal support of many frames. In the approach described here, complex sinusoids are used instead of polynomials and the designer is free to choose both the extent (i.e.…”