Invariant Features Extraction Of Typical Targets

Invariant features extraction of typical targets is a key issue in the field of targets recognition. In view of the factors such as moving invariance, rotating invariance and scale invariance owned by images, an algorithm of integral projecting based on circular co-ordinates is proposed, which is name as circular integral projecting(CIP). With the alternation of integral mathematic objects, the curves corresponding to the circular integral projecting describe different texture information in image. The value of curve must be normalized. It's found the curve's phase just moved while the same target rotated, and after expanded curves corresponding to different rotating angle, there were the same waveform in each period of different curves.According to the variation of the same target regions or the same class target regions, we define the two conceptions: strict constrain feature curve and loosen constrain feature curve. In the same target regions and the same class target regions, homogeneity and difference exist at the same time in the details of texture and texture distribution. If the details in the two different images are same with each other perfectly, curves corresponding to the two images become a strict constrain feature curve pair. While homogeneity decreases and difference increases certainly, curves become a loosen constrain feature curve pair. Feature curves describe the homogenous and different information in the typical target regions. Invariant feature extraction is equal to extract the homogenous information. Based on the simple prior knowledge, an approach named as multi-scale low-frequency curve analysis based on standard orthonormal wavelets is proposed. Two reconstruction formulas respectively base on wavelets and low-frequency functions are used to reduce computing burden. This article proves that under the conditions of standard orthonormal wavelets, the two formulas are equal with each other. The formula based on wavelets is used to decompose feature curve function quickly; then the formula based on low-frequency functions is used in multi-scale analysis of .In project experiment, a cluster algorithm based on simple prior knowledge and connective region method is designed to decrease computing burden compared to cluster algorithms discussed in this article. A region expansion method based on sobel region sign and morphological dilation is proposed which is equivalent to traditional expansion method. But the new can overcome the default of the traditional one efficiently where the judging criteria are invalid. At last realize the analysis and segmentation of the typical targets regions. And apply the conclusion in the research of matching images obtained by different sensors.