Research On Detection And Recognition Techniques For Specific Targets

Image is an important medium for human to acquire information.Compared with the signals such as text and sound, image contains moreinformation. It’s very high in cost to explore useful information from hugeimage data, so it’s of great significance to develop machine vision so as toliberate the human visual system. Automatic object detection andrecognition is an important branch of computer vision. Object detectiontechnology has been applied in many fields, such as industrial production,intelligent transportation, video surveillance, semiconductor manufacturing,medical diagnostics and intelligent navigation. We can list a great numberof successful applications of object detection techniques, however, objectdetection technology is still in development. In fact, object detectionalgorithm is still an open problem and its theoretical framework is still inthe exploration. As for the application of object detection technology in areal system, real-time capability and robustness are two big challengesfacing the object detection system. Especially in complex scenario, it’s stilla pending issue.The main task of this thesis is to explore the detection and recognitiontechniques of artificial targets in complex scenario. Generally, artificialobjects present distinct texture, color and shape features during imagingprocedure. In complex scenes, however, such features are destroyed tosome extent due to imaging degeneration. For example, in underwaterenvironment, images are often subject to color cast problems. In heavilynoisy environment, image edge is easy to be damaged. Focusing on these issues, this thesis aims to provide an effective and robust object detectionalgorithm framework in complex scenario. The major contributions in thiswork are listed as follows:For the heavy noise and severe color cast problems of underwaterimages, we proposed a color cast correction algorithm based on thegeneralized image equalization model, which can partially restore the colorfeatures of a degraded image. Then we applied mean-shift filter to suppressthe image noise. These pre-processing steps provide a guarantee for thesubsequent object detection and recognition.We proposed an image transform technique based on color channelratio features and then we developed a region of interest (ROI) detectionalgorithm. The color channel ratio image (CCRI) is illumination invariant,thus effectively suppressing noise and removing shadows. CombiningCCRI and k-means algorithm together, we can easily implement a fast andefficient ROI detection method. Meanwhile, based on contour extractionmethod and flood-fill algorithm, we proposed an ROI optimizationalgorithm, which can effectively remove false detection area.In order to recognize and localize objects in scene, we utilize a threedimensional distance transform and directional chamfer matchingalgorithm to match the object with given templates. The chamfer matchingalgorithm is scale invariant and insensitive to slight shape deformations. Inorder to improve the algorithm performance,3D distance transform,segment representation, distance integral image and optimal region searchtechnique are introduced to optimize the shape matching algorithm. Inaddition, based on the algorithms above, we develop a system withautomatic object detection, recognition and tracking functions on C++andOpenCV platforms.