Research On Image Matching And Object Location Based On Multi-Scale Feature

The complexity of the underwater environment results in the problem of lacking information underwater,which lead to the high difficulty in dealing with object recognition and localization. The situation becomes even worse when we wish to extract the features of non-structural object and recognize it.The multi-scale representation has attracted considerable attention recently since it describes the the performance of the real world. At present, there are a variety of explanations about the concept of scale such as resolution, image size, quantitative, the distance and convolution kernel and so on. This dissertation aims at the seafloor object and undertakes some works such as constructing the multi-scale model, multi-scale feature extraction, the image matching , the application of object localization and so on.The research is based on Gaussian scale space in the underwater environment.Firstly, the development and present situation of vision-based multi-scale theory underwater were reviewed, and the definition of multi-scale and its basic theory are discussed. Then the characteristics of different scale model are introduced. We focus on the Gaussian scale space theory. Gaussian scale-space is an important domain in image study. Because of the simplicity of the theory and the unique property of Gaussian function, Gaussian scale-space is applied in scale-space theory. As the scale increases, more and more information is lost. Those information can be preserved by DOG function (Difference of Gaussian). It is crucial to choose appropriate scale parameter during constructing Gaussian scale-space.At present, Many applications in Gaussian scale-space about the scale parameter is not clear, which may lead to bad effect of layer. The paper proposes a kind of adaptive algorithm of scale parameter in terms of the module of visual characters. Experimental results show that the uniformly distributed information in scale-space will be useful for higher-level image processing technologies such as object recognition.Scale invariance is a key factor to evaluate the feature extraction algorithm. For the noise of underwater image, a kind of scale-invariant feature extraction algorithm is proposed which hold the anti-noise characteristics and based on Harris comer and Gaussian scale space.We extract the Harris corners in different scale. Then, the points detected at the highest level of the pyramid are correctly propagated to the lower level. The corners detected repeatedly in different levels are chosen as final feature points. At last, the characteristic scale is obtained based on maximum entropy method. The experimental results show that the algorithm has low computation cost, strong anti-noise capability and excellent performance in the presence of significant scale changes. Due to the invariance of scale,rotation,illumination,SIFT (Scale Invariant Feature Transform) descriptor was commonly used in image matching domain. Feature extraction experiments are carried out with SIFT.Conbined with traditional Entropy a concept of Multi-Scale Entropy was given. The non-structural object contours of the hydrothermal chminey undersea are extracted with the advantages of the Maximal Multi-Scale Entropy Difference (MMSED), which provides the basis for the following processing.We next mainly discuss the matching methods in underwater environment based on SIFT. In order to increase the number of matches,color feature extraction is analyzed. The potential value of the method in the color image is presented,which can provide reference to the future work. However, the fact the presentation of one feature point needs 128 dimensions that will reduce the algorithm efficiency of real-time.we develope a simplified algorithm based on SIFT (SSIFT) to express a feature point with only 12 dimensions which based on a circular window to improve the efficiency of matching. The experiment results show that the algorithm can reduce the rate of time complexity and maintaining a relatively good matching rate at the same time.Furthermore, there is no quantization errors in SSfFT which exist in SIFT resulted by orientation assignment.Finally, we discuss the object localization application underwater. We describe the the relationship between the different coordinates as well as the methods of camera calibration. The plane camera calibration algorithm of Zhang's is then used to get camera internal calibration. The fundermental matrix and essential matrix is obtained by SIFT matching. After the decomposition of external motion parameters with single-eye camera, we use structure from motion method to obtain three-dimensional information and carry out object localization. Finally, we give the three-dimensional information acquisition process and the computer simulation result.The results show that the method can meet the demand of underwater manipulation precision.A series of studies have been done in this dissertation for the object localization based on underwater robot manipulation. The conclusion can provide theory basis and technical means for the underwater intelligent robot.