Research On MSER Local Invariant Feature Extraction Algorithm In Multichannel Images

The changes of viewpoints, scale and rotation and the generalized affine transformation of blur, local block and complex background always exist between most relative images. However, in many image processing and applications these changes must be removed. Consequently, a stable and repeatable detecting feature is need. Local invariant feature as a feature with ability of stableness, repeatability and matching score is been researched intensely.Maximally Stable Extremal Region (MSER) has been proved to be a powerful local invariant feature. It is been used in many image applications such as wide baseline matching, 3D reconstruction and image segment. The traditional MSER only utilizes the intensity information to extract invariant features, but does not use the color information of every channel.According to the human vision sense of color, the HIS color space is more nature, intuition and consistent with human vision characteristics than other color spaces.,so the extraction result of local invariant feature based on HIS could descript the nature of image. Also the independence of three channels of HIS can reduce the complexity of color image processing and the processing time. So we extract MSER features from the three channels of HIS to take full use of the information of three channels,However, because the stableness of features from different channels is different, especially when the imaging conditions are changed, the stableness of features from simple MSER feature extraction method in three channels is changed.To solve the problem, this paper presents an improved MSER feature extraction method in three channels. This method firstly utilizes the SVM to train the stable characteristics of MSERs with good stableness and to get classification method with different accuracy for each channel. Then the MSER features of each channel are filtered with an automatic accuracy selection filter method which avoids the depression of feature quality because of the unstableness of features. Also we compare our method with traditional MSER method and simple three channel MSER method in the conditions of existing generalized transformations of viewpoints, illumination and blur etc between images. Experimental results demonstrate that the proposed method outperforms the traditional MSER detector.