Research On Image Local Invariant Feature Extraction

Image local invariant feature extraction is a fundamental problem in the field of computervision. The local invariant feature extraction is roughly divided into two steps: firstly,detect thefeature point position,scale and other related information on the image by local invariant featuredetection,then describe the feature points based on the detected information and neighborhoodinformation. The core issue of the image local invariant feature extraction is to improve thespeed and accuracy of the feature extraction, as well as to strengthen the robustness and stabilityof algorithm,which makes the local invariant feature extraction techniques satisfy the demand ofreal-time applications.This paper focuses on extraction algorithm based on the image localinvariant feature, which mainly studys from the following two aspects:Firstly,the corner detection algorithm of low complexity adaptive window is proposed toimprove the detection efficiency of the Harris operator. At first, average gradient amplitudes arepreprocessed to remove most unrelated points whose changes in the brightness are smaller, andthen only retained interest points are operated by Gaussian filter and response function. If theinterest points whose response function value are more than a certain threshold are identified ascandidate corner points.Finally, the candidate points is dealt with by non-maxima suppression ofadaptive window,no other candidates corner in the neighborhood window is necessary to dealwith by non-maxima suppression of adaptive window if the candidate corner is the finalcorner.After the improvement of this paper, the algorithm saves computation time and improvesspeed of corner detection.Secondly,combining with character of image registration field, this paper introduces themulti-scale analysis and related cluster analysis to design a local invariant feature extractionalgorithm based on related cluster analysis in order to improve the speed and accuracy of imageregistration. The algorithm firstly calculates the overlap region of the largest rectangle in thesmall-scale image testing and matching feature points,then clusters the overlap region of fourboundary points as the cluster center in the overlap region into four classes, finds out the most relevant class of matching points as a reference image and the unregistration image region, at lastonly detects feature points of the relevant area in large-scale image, and removes false matchingand calculation of the geometric transformation parameters by the modified RANSAC algorithm.Simulation result shows the designed in this local invariant feature extraction algorithm can notonly improves the speed of image registration, but also accurately calculates the transformationparameters between the images.Finally,I summarize the local invariant feature extraction algorithm of this paper andprospect the future work and intensive study.