Research On Partially Occluded And Broken Object Recognition Method Under Affine Transform

Object recognition is always the hot topic in computer vision. The distortion andthe partial occlusion and break in the image are the conmen problem and also theclassic difficult one in object recognition. Therefore, the research based on theseproblems is very significant. And the chosen for a reliable description method andmatching one is the most important problem in it. In many applications in patternrecognition, if the boundary of the object contains enough information, then we cansimplify the object recognition as the planar curve matching problem.We mainly do research in the description and matching method in contour curveof the object in image when distorted or there exists partial occlusion and break in thecontour of the object.Through the theoretical analysis and experiment comparison, we choose thecanny edge detection operator to extract the object contour. In order to remove thenoise, we use Gaussian smooth function to filter the image to get a smooth contourcurve of the object. Then we can get the ordered sequence of the coordinate of thepixels on contour curve by boundary tracking. Finally, we use an improved cornerdetection algorithm based on CSS corner detection to detect the corners in ourexperiment images and obtain a better detection result.For the first time, we define and extract a new affine invariant point——chordheight point. The feature points are sampled from each sub-curve and they candescribe the curve more precisely than the conmen feature points (such as the corner,inflection point and point of tangency). And they can solve the problem that the curvecan’t be described accurately when the curve is smooth. This method is suitable fordeal with the image under affine transform due to its affine invariant properties andthe invariant vectors constructed with these feature points are also affine invariant.This feature points are robust to partial occlusion in the contour because the chordheight point and the new constructed affine invariants are both the local invariance.Otherwise, we add the concave-convex judgment for the sub-curve based on matchingthe recognition vectors. So we can also match the curve of different feature vectors,but also tell the sub-curves of different concave-convex property apart. Theoretical analysis and experiment result demonstrate our algorithm is right.We improve the existed SVD (singular value decomposition) algorithm appliedfor matching curve by using the properties that there only exists rotation relation afterSVD to curves in the subspace. Firstly, we construct an objective function andcalculate its minimum to find the optimum matching segment as the initial matchingsegment. Then, along the curve at the counter-clockwise direction and the clockwiseone, we search the other matching parts of the curve which are satisfied with thematching requirement by extending the initial matching segment. Theoretical analysisand experiment result demonstrate that the improved algorithm can match the curvewhich there exist partial occlusion or break and we obtain good matching result.Our proposed algorithms in this paper can deal with the image under affinetransform and partially occluded.