Machine Vision, Object Recognition Research And Explore

Computer Vision aims at enabling and advancing intelligent perception of input image data.Image content is understood by recognizing objects in images;thus object detection/recognition is a very fundamental subject.An efficient object detection algorithm is a basis and prerequisite in many applications,including but not limited to:content-based image retrieval,video surveillance,medical image processing,industrial robots.Object recognition techniques will influence industry, medical science,traffic control,national defense,and possibly change the traditional way in which they are currently developing.The wide-spreading of this technique is expected to become an integral part of our daily lives.However,despite its fast-developing pace,object recognition is still at its first stage.There are many detailed algorithms designed for specific applications,although no general yet robust algorithm framework is available.This paper presents what I have done on this research topic.In Chapter 3,a simple but efficient object recognition algorithm is introduced based on basic point-shape-feature.In this algorithm,improved Shape Context features are used to find matches in images,followed by a generalized Hough Transform to organize qualified matches.Object hypotheses are selected from this voting procedure.The Shape Context is improved mainly to avoid background clutter and gain tolerance to shape deformation.This kind of finding matches using pre-defined object model is called top-down recognition.It usually has a high recall but suffers from low precision.To overcome this,hypotheses are further verified using a discriminative classifier.Object regions in image are finally obtained by combining the cues from bottom-up image segmentations.Background clutter is always "annoying" to the performance of object recognition.It often corrupts image features by adding useless information to feature dimensions.Most pervious methods learn which dimensions are more important using many training exemplars.Chapter 4 proposes an object detection method in heavy cluttered background using only a single exemplar.It starts with bottom-up contour grouping as its basic perception elements,and uses a much larger context for shape feature extraction.Normally,enlarging the context area will worsen the problem of background clutter,especially for those objects with elongated structures. We tackle this problem by introducing a selection process on image contours,and model contours as well,during which contour integrity is maintained.Selected image contours are compared against selected model contours.The experiment results show this selected shape feature can remove background clutter effectively,and achieve good detection results.Object recognition is not only about telling the existence or positions of objects inside an image,but also about conducting a level up analysis,including pose estimation,etc.Chapter 5 introduces a pose estimation algorithm utilizing the same SELECTION strategy.Selection is not merely on image contours,but on model pose parameters as well.Valid poses are proposed by matching features generated by selected image contours against different model poses.This algorithm demonstrates its performance on a baseball image database.Chapter 6 starts an important sub-topic of object recognition:contour clustering. Unlike top-down recognition using contour selection,which can also be seen as a contour clustering process using pre-defined model knowledge,herein contour clustering is still a bottom-up process via another related image(stereo,motion, similarity).Street scene is an important play stage for vision algorithms.Besides pedestrian, there are lots of other interesting objects on the streets.We are particularly interested on objects like {traffic lights,road signs,lamps,fire hydrant,trees and car} and we want detect them on the streets.Due to the inter-object variations(some are rigid objects,some are composed of texture region,some are semi-rigid or deformable objects),hybrid detectors are exploited to detect them efficiently, detailed in Chapter 7.Shape feature is mostly extracted from an edge map.To acquire a clear edge map thus improve shape feature descriptor,in Chapter 8,a structure-preserving image diffusion technique is proposed using adaptive grouping-based bandwidth selection.The goal is to preserve image structure while remove noises,especially random and texture noises.The contribution includes a novel diffusion kernel which has better structure preserving performance,and an adaptive selection on kernel parameters.By experiments,we find it can improve shape feature descriptors by enhancing the edge detection results.