Research On Optical Images Fast Detection Method For Multiclass Target

With the development of information technology,a large number of optical images can be obtain through video,remote sensing and other means,how to find useful information from massive optical images is the hotspot of current research,target detection is its core research content.The target detection includes single frame static image detection and multi-frame motion image detection.In this thesis,we proposed a series of multi-class target detection methods for optical images based on a number of static image detection projects developed by the research group.Mainly include the following five aspects:(1)A Logarithm-Based Image Denoising Method for a Mixture of Gaussian White Noise and Signal Dependent NoiseDue to influence of internal or external noise,some noise produced in the image,In existing literature,the majority of denoising methods deals with Gaussian white noise,there exists fixed pattern noise and a mixture of Gaussian white noise and multiplicative noise in the actual,In order to obtain a more intuitive picture and ensure the stability of subsequent feature extraction,this thesis proposed a novel method for reducing noise in an image corrupted by a mixture of Gaussian white noise and signal dependent noise.In this method,the multiplicative noise is transformed into additive noise by the form of logarithmic transformation,and the additive noise is eliminated by the classical additive noise image denoising method.Finally,transform this noise-reduced logarithm image back to the original space.Experiments verify the feasibility of the proposed method.(2)A method based on shape contour descriptionThe existing method to describe the shape of the contour have some disadvantages,such as,contour feature extraction and calculation need a large amount of calculation,the efficiency is low.we present a kind of Fourier descriptor based on shape matrix,This thesis describes the generation of the shape matrix,the construction of the Fourier descriptors based on the shape matrix and the similarity measurement.(3)An image detection method based on SVM and active learningTo further improve the detection speed and detection precision,this thesis proposed a combination method of SVM and active learning target classification,the feature extraction based on shape Fourier descriptor of the matrix,the characteristics of the sample data in the database divided into two stages,the first phase to obtain representative data,as the initial labeled data,to reduce the scope of testing.The second phase access to key data,the key data as the most valuable label samples,through active learning,constantly iteration,until the end of the meet the conditions.(4)A quick detection method for remote sensing image based on regional target linear modelBecause the ground object of interest in the remote sensing image usually contains a large number of straight lines features,and the straight lines feature is stable with respect to the point feature,it is simple to compare with the contour feature,so the research of the method is proposed.Firstly,the model marked the target of the region from the historical satellite image,then,a large number of straight lines extracted from the data to detect.Finally,the regional target detection and positioning achieved by matching the two straight sets.Using the voting mechanism of the offset accumulator to achieve the goal of fast and accurate positioning,offset the maximum amount of compensation is the best match position,including a variety of targets,the target has been a large range of occlusion,small angle rotation of three scenes on the satellite Remote sensing images were tested.(5)Segmented Three-order Tucker Decomposition for Hyperspectral Image Anomaly DetectionAn improved anomaly detection method for hyperspectral images based on three-order tensor decomposition is proposed.Compared with existing methods,the method does not require large memory and strong data correlation.The large three-order tensor divided into a number of small sub tensors to realize the stable detection of abnormal targets.The main process is as follows: the hyperspectral data cube is represented by three order tensor;Segmentation of hyperspectral image data in spectral and spatial domain;The Tucker decomposition is performed on each sub tensor;Reconstruction of tensor for anomaly detection;Anomaly detection is performed in each sub tensor and the detection results are fused.By comparing with some popular detection methods in synthetic data and measured data,the proposed algorithm achieves better detection performance.