Research On Segmentation And Recognition Algorithm Of Multi-source Image

Image segmentation is one of basic problems in computer vision, which always opens image processing. After several decades of development, the field of image segmentation has matured. Results obtained using different algorithms differ from each other. This is not only caused by the imprecision of segmentation algorithm, but also due to the various levels of information in the segmentation results. It is challenging to segment images accurately and efficiently using these information.In this paper, we summarize several image segmentation methods and the features of their results. We propose a novel framework for automatic image segmentation. In this approach, a mixture of several over-segmentation methods are used to produce superpixels and then aggregation is achieved using a cluster ensemble method. Generated by different existing segmentation algorithms, superpixels can describe the manifold patterns of a natural image such as color space, smoothness and texture. We use them as the initial superpixels. Grouping cues which affect the performance of segmentation can also be captured. After the over-segmentation, the simultaneous collection of superpixels is expected to achieve synergistic effects and ensure the accuracy of the segmentation. For this purpose, cluster ensemble methods are used to process the initial segmentation results and produce the final result. Compared to state-of-the-art techniques, we have achieved competitive results on the Berkeley Segmentation Database. The performance is quantified using four criteria: PRI, BDE, VoI and GCE. Using the proposed framework, PRI rises from 0.7735 to 0.8059, and BDE reduces from 13.3087 to 12.0407.Targets detection and identification start with the results of segmentation. In this paper, we focus on targets detection under the specific scenes, and proposed two infrared identification algorithms. According to the characteristics of the scene, we firstly pick foreground objects using different segmentation methods, and then filter those objects. And finally we get object recognition results using shape matching. Compared with other detection algorithm, our method is more robust.