Algorithms Of Image Feature Extraction Based On Visual Information

Computer vision is a popular approach for two-dimensional image recognitionand analysis. It is a simulation of human visual system and it takes visual informationas its proceeding object. Visual information includes many image features such asshape, location, color, texture and so on. The main task which computer vision dealingwith these image features is feature extraction, we hope it can preserves its invarianceas far as possible. Texture and contour are studied in this paper, and we also improvesSIFT algorithms of feature points.The single object contour extraction algorithm according to texture feature in thecomplex background is easier and faster than the conventional one which edgeextraction make convolution between the proper operating with reference image basedon the space field. But there are several shortcomings such as loss of some contextinformation, unable to divide object with background, discontinuous and too narrowedges. Contour is a middle feature which includes shape and structure, we canrecognize it and distinguish background and object. Advanced visual pathway consistsof what pathway which perceived shape and where pathway which perceived location.After Hubel and Wiesel introduced the method of classified processing, we began tostudy the structure of abdominal lateral flow. Serre gives the HMAX model which wascombined with biology in2007. The HMAX model is a hierarchical structure consistS1layer(simple cell layer), C1layer(complex cell layer), S2layer(synthesis of featurelayer), C2layer(complex feature layer), VTU layer(visual adjustment cell layer).HMAX model is similar to the way of identification by human. Obviously, it prefersto identify the object as it export eigenvector at the final VTU layer. But HMAXmodel cannot obtains the final figure according to this feature extraction. As the effectis not deal to recognize background edge and object contour by traditional edgeextraction algorithm, we weaken the background edge by non-classical receptive fieldrestrained neurons. Therefore, on-classical receptive field can improve quantity ofcontour extraction but weaken edge control. There is great difference in structure ofvisual of cerebral cortex of human being. Thus to get better effect of contour extraction, we structure a new model consists of five layers and its hierarchy divisionconsistent with HMAX model. Our algorithm set parameter every layer to improveHMAX model. We verified the validity of the new algorithm by evaluating its effectdrawing, accuracy and box-and-whisker box in simulation experiment.LBP face algorithm recognition according to contour feature extraction islow-resolution. LBP operator, which is based on gray scale, can extract contexturefeature from a figure. It is more simple to implement and easier to understand. It alsohas strongly robust when it brings changes due to gray scale. Moreover, it is powerfulin contexture classification. Multi-resolution is important in wavelet analysis. It putsup a frame for wavelet structure in view of functional space. Every figure showsdifferent feature with different resolution, it looks more clear and the detail is moreapparent with high-resolution. We focus on multi-resolution to deal with the figure.Researchers work on multi-direction study after Mallat present multi-resolution theoryin1987. Wavelet transform is based on Fourier transform. Fourier transform is aglobal transform while wavelet transform is a local transform. They both aretime-frequency analysis method, but wavelet transform belongs to transform of timeand frequency, space and frequency which involve useful information which can becaptured. Wavelet transform makes epoch significance in harmonic analysis. Thispaper present LRLBP algorithm combined with LBP algorithm which has strongrobust and low-resolution and wavelet transform which has filtering function. Thisalgorithm takes face recognition for object, begins with low-resolution figure,decomposes figures by proper wavelet transform and extracts feature vector bycombining with LBP algorithm. It has better effect and easy calculation. We test thisalgorithm by extracting figure randomly from ORL database and YALE database, andtest result manifests its whole performance is better than basic LBP algorithm.The definition of invariance was presented in the early19century according tosimplified SIFT algorithm which extract feature points. As study deep, researchersdiscovered the visual process that human visual system selecting the invariance issimilarly with the visual invariance theory. Naturally, researchers introducedinvariance theory into machine visual, such that machine visual physiology steeply.Algebraic invariance focus on global general view and it has strong anti-noise performance, while differential invariance focus on local detail and it is not influencedby location shift and towards change. SIFT algorithm, published by D. G. Lowe in1999and summarized in2004, is based on extracting local feature. It can calculatemassive data and jointly match with other form of feature vector. The main problem ofSIFT algorithm is its too high dimension, and some algorithms tried to obtain betterresult by decreasing its dimension. This paper presents an S-SIFT algorithm based onthe ideal of decreasing dimension and improves SIFT algorithm. S-SIFT algorithmchanges the shape of feature area and rectifies rectangle in SIFT algorithm into circleand preserves its range. These changes of neighborhood shape avoid Gauss fuzzyprocess and without adjustment of coordinate axes. The dimension of feature vector ismuch decreased and its towards, location and scalar transform is invariant in S-SIFT.We test its dimension, invariance and time complexity by simulation experiments, andthe result shows S-SIFT algorithm is not as good as SIFT algorithm in match effectbut it is better in synthesis performance, its time complexity is far less than SIFTalgorithm.