Research On Image Feature Extraction And Its Application

High data dimension and increasing complexity of data type occur in computer vision and image processing field. The cost of typical calculation and analysis of such image data is too high. Sometimes, the typical calculation and analysis even fail completely. Generally speaking, feature extraction of the sample data should be finished before the analysis and processing of the high complex image data. The process can including extracting feature points which are most related to the samples, removing the noisy feature points and the redundancy feature points that are not related to the sample data, then the valuable data for the next process can be obtained. And the image processing result is depended on the quality of features which are got through the image feature extraction. Image feature extraction acts a key role in fast image registration, target recognition, image retrieval and so on. Therefore, image feature extraction methods with well-expressive and noise resistance ability from image data are still difficult problems in image processing field. This thesis focuses on the study of the image extraction in many applied domain-specifics, and comes up with a series of methods to improve the efficiency and feasibility of the extraction on the basis of deep analysis of the usual extractions. The main contributions are listed as follows:1 The extraction of nighttime images with time frequency weighted and constrained optimization evolutionary method is proposed. The treat of multi-frame nighttime images in time domain and frequency domain and weighted processing as the same time are an advanced method for deep extraction of night blurred image. In the traditional processing method, time domain is extracted only and cannot be used to analyze and extract the spectral characteristic of the multi-frame night blurred image. The extraction of nighttime images with time frequency weighted and constrained optimization evolutionary method is proposed in this thesis. For the row image, first the related information between multi-frame images in frequency domain and time domain is extracted. Then the new image features are formed with the weighted processing. Finally, the loop optimization is made continuously on the results of the feature extraction through restraining optimization method to achieve better results.2 A feature extraction of noisy image is raised on the foundation of bar chart feature description. In the content-based image retrieval method, the image content usually is expressed by means of bar chart, and the retrieval process is made in accordance with similarity of the bar chart. However, the bar chart becomes smooth and image becomes more similar due to the noisy obtained in the image. As a result, the number of images'return increase with low accuracy rate. This thesis proposes a bar chart feature descriptor which is not sensible to noise and is used in image retrieval. First, the noise in the image is described as steady additive Gaussian white noise, and then a related bar chart is drawn. Second, feature descriptor is defined by the variable origin moment, and an analysis is made about how to restore the bar chart of row image through the feature descriptor.3 An extract method which is based on aviation image characteristics of LBP (Local Binary Pattern) and a human visual perception model is proposed in this thesis and the retrieval problems of the aviation image are studied accurately. Aviation image target distribution is random in a particular area, and jamming targets interferes with the goal and identifies the target distribution. Traditional ideas are based on the goal of image retrieval method, according to the characteristics of the specific target pixel. Interference ruled out with a characteristic contrast is given priority to, more machinery, the goals of environment, and the eliminating process is extremely complicated and low efficiency. Based on human visual perception, a kind of air target image retrieval method put forward to avoid the above defects. Local binary method for the image feature extraction and the characteristics of different kinds of sense as the human eye visual space data are applied as the basis of image retrieval. Then the efficiency and accuracy of retrieval method can be improved by the human visual perception model.4 A new method based on SIFT(Scale-invariant feature transform) descriptor is proposed to solve the problem of image registration. In this algorithm, this paper uses the nearest distance ratio method to obtain the initial matching feature descriptors to reduce the effect of outliers. In addition, the use of the scale direction of the joint restriction is applied to search for the false SIFT descriptor pair. By using samples randomly to delete the abnormal value, the accuracy of image registration is improved.5 In order to achieve the real-time edge detection of image, a real-time edge detection system based on embedded multi DSP is studied. The image edge detection module has three DSP, and each DSP is connected by two other. The image is pretreated by FPGA to improve the detection efficiency and to ensure the instantaneity of the edge detection. The detection algorithm is the improved universal gravitational edge detection algorithm. The algorithm can reduce the impact of noise for the detection result, and increase the detection accuracy of details.