Sub-pixel Estimation Methods Research Articles

Discovery, Quantitative Recurrence, and Inhibition of Motion-Blur Hysteresis Phenomenon in Visual Tracking Displacement Detection.

Motion blur is common in video tracking and detection, and severe motion blur can lead to failure in tracking and detection. In this work, a motion-blur hysteresis phenomenon (MBHP) was discovered, which has an impact on tracking and detection accuracy as well as image annotation. In order to accurately quantify MBHP, this paper proposes a motion-blur dataset construction method based on a motion-blur operator (MBO) generation method and self-similar object images, and designs APSF, a MBO generation method. The optimized sub-pixel estimation method of the point spread function (SPEPSF) is used to demonstrate the accuracy and robustness of the APSF method, showing the maximum error (ME) of APSF to be smaller than others (reduced by 86%, when motion-blur length > 20, motion-blur angle = 0), and the mean square error (MSE) of APSF to be smaller than others (reduced by 65.67% when motion-blur angle = 0). A fast image matching method based on a fast correlation response coefficient (FAST-PCC) and improved KCF were used with the motion-blur dataset to quantify MBHP. The results show that MBHP exists significantly when the motion blur changes and the error caused by MBHP is close to half of the difference of the motion-blur length between two consecutive frames. A general flow chart of visual tracking displacement detection with error compensation for MBHP was designed, and three methods for calculating compensation values were proposed: compensation values based on inter-frame displacement estimation error, SPEPSF, and no-reference image quality assessment (NR-IQA) indicators. Additionally, the implementation experiments showed that this error can be reduced by more than 96%.

10312 画像相関法におけるサブピクセル推定手法の比較と評価

10312 画像相関法におけるサブピクセル推定手法の比較と評価

Subpixel Urban Land Cover Estimation

Three machine learning subpixel estimation methods (Cubist, Random Forests, and support vector regression) were applied to estimate urban cover. Urban forest canopy cover and impervious surface cover were estimated from Landsat-7 ETMimagery using a higher resolution cover map resampled to 30 m as training and reference data. Three different band combinations (reflectance, tasseled cap, and both reflectance and tasseled cap plus thermal) were compared for their effectiveness with each of the methods. Thirty different training site number and size combinations were also tested. Support vector regression on the tasseled cap bands was found to be the best estimator for urban forest canopy cover, while Cubist performed best using the reflectance plus tasseled cap band combination when predicting impervious surface cover. More training data partitioned in many small training sites generally produces better estimation results.

Quantitative Evaluation of Image Interpolation Methods Supposed to Be Applied to Sub-Pixel Estimation

Sub-pixel estimation of the corresponding positions in area-based matching is often conducted to obtain the precise locations of objects using stereo images. Such sub-pixel estimation methods as least squares matching and cross correlation method using sub-pixel shifted images require appropriate interpolation of the gray values of the window corresponding to the template. Since there are few reports on comparison of image interpolation methods in sub-pixel estimation accuracy, we decided to investigate performance of interpolation methods applied to sub-pixel estimation. This paper reports the preparatory experiment conducted in order to evaluate image interpolation methods quantitatively by using 54 diverse images. Three popular methods in remote sensing and digital photogrammetry: bi-linear interpolation (BL), bi-cubic interpolation (BC), and cubic convolution (CC) were investigated. The experiment results demonstrate that the interpolation accuracy of all three methods is correlated to texture measures of the image such as grey level difference vector measures and spatial frequency. Furthermore, the experiment results show that BC and CC can produce better interpolation results than BL, when an image has no noise or smaller noises. On the other hand, BL and BC can produce better interpolation results than CC, when an image has larger noises.

テンプレートマッチングにおけるサブピクセル推定法の精度比較

Precise results of matching are required to obtain the precise locations of objects using stereo images. Accordingly, sub-pixel estimation of the corresponding position in area-based matching is often conducted. Several methods to obtain sub-pixel estimation in area-based matching have been proposed. However, there are few reports on accuracy comparison of those methods. Therefore, we performed an experiment in order to investigate accuracy of sub-pixel estimation methods in area-based matching by using 54 diverse images. Our study focused on two popular methods in digital photogrammetry: cross correlation method with fitting a second order surface by using normalized cross correlation coefficients (CMF), and least squares matching (LSM) . The experiment results demonstrate that the accuracy of sub-pixel estimation methods in area-based matching varies considerably with the properties of the image. Furthermore, the experiment results show that LSM can estimate the corresponding position more accurately than CMF. The estimation error of LSM would be average half of that of CMF.

Use of SPOT4-VEGETATION satellite data to improve pasture production simulated by STICS included in the ISOP French system

In France, pastures represent a significant land-cover type, which mainly sustains husbandry production. For this reason, it is of great benefit to develop real-time monitoring of pasture biomass production, taking into account its spatial and temporal variability. The absence of low-cost methods applicable to large regions has oriented French stakeholders to the use of growth simulation models adequately informed through spatialised databases (such as the ISOP system). Remote-sensing data may be considered a potential tool to improve simulations by objective observations in a real-time framework and the aim of this work was to evaluate this potential role of remote sensing. Thirteen forage regions (administrative partitioning of the French territory for pastures and grasslands) were selected in France, differing by their soil, climatic and land-cover characteristics. SPOT 4 -VEGETATION satellite images (1 km 2 resolution) were used to provide the spectral signature corresponding to pure pasture, using subpixel estimation methods. This information was then related to growth variables calculated by the STICS-pasture model (included in the ISOP system). We found that the best relations were obtained between a middle infrared-based vegetation index (SWVI) calculated from the elementary reflectance bands of the satellite, and the leaf area index (LAI) calculated from STICS. The use of these relations first showed the ability of satellite data to provide real-time estimations of growth status variables. Second, the comparison between both types of data showed that spatial and temporal differences existed between satellite and model information, mainly during the harvesting periods. This result could contribute to improving the model evaluations on a regional scale.