Reconfigurable Computing Technology Based Research On Image Recognition And Classification System

Along with the rapid expansion of application areas of image automatic target recognition (ATR) and classification technology, the demand of computer processing abilities is getting higher. Reconfigurable computing is a brand-new technology which is developing with the progress of electronic technology. It combines microprocessor's flexibility and ASIC's efficiency, and makes the realization of real-time image processing technology rapidly and possibly.Based on comprehensive analysis of current image recognition and classification algorithms'characteristics, deeply investigation of SAR image's behaviors, target's features and extraction methods, this thesis has done the beneficial researches and discussions to the usage of reconfigurable computing technology to realize the algorithms of image recognition and classification. The main tasks and the innovations of this thesis include:(1) A study of traits of remote sensing image and algorithms of target recognition was made, a feature templates matching based ATR algorithm is proposed.Because of wide coverage, in-time and comprehensive data etc., remote sensing image becomes an important data source for every branch of national economy and remote sensing imagery processing turns into hotspot in field of computer-based applications at present. The ground objects of SAR imagery can be divided into point target, line target, area target and hard target that combine by them in limited scales. Most of artificial targets are hard targets. The hard target that displayed in SAR image is not so similar to ordinary sensation visual image. The details of the targets reveal if and only if the radar resolution is high enough, so the recognition preferably when the targets'features are sufficiently analyzed. After analysis SAR ATR process and in the light of domestic and foreign research experience, a templates matching based ATR algorithm is proposed. Via effectual pretreatment of goal image, the targets limited in regions that content less background. This method can not only decrease the background's influence to target detection but also reduce the data size. Though analyze to the features of the targets, workable templates are gained and recognition rate is enhanced.(2) Research on pattern noise, textural property and support vector machine, a texture based image classification algorithm is proposed.Image classification is a kind of procedure that according to the pixels'grey scale and space characterization to justify the category of ground object. In application of sensing image, image classification is demanded in specialized information extraction, motion prediction, making thematic map, establishment remote sensing database and so on. Depending on whether the priori knowledge is utilized in discriminant function's solving, the classification methods can be referred to supervised and unsupervised. Unsupervised methods can only find differing aspects among the samples but cannot fix the property, so the accuracy and velocity of convergence are relatively poor. This thesis focuses on supervised methods. After discussion the insufficient of available method, SVM is chose as classify. Though effectively denoise and extraction the textural property, the computation of the irrelevant eigenvector is reduced and the rate of correct is improved.(3) Investigate the methodology of hardware/software co-design and apply the module-based reconfiguration design method to implement the image recognition.In terms of large data size and simple realization characters in low level of recognition of high-resolution image, a reconfigurable platform based hardware/software co-design method is proposed. Analysis the proposed image recognition algorithm in connection with the characteristics of reconfigurable platform in the first place. And then designs each module by using modular design methodology. Finally implements the run-time partial reconfiguration in actual devices. The design documents are not only universally valid in Virtex series reconfigurable logical component, but also can be used in pretreatment all sources of image and have laid a good foundation for realizing other interrelated image processing algorithms. Results show that the system can improve the recognition rate significantly, have better speed performance when compared to a general purpose processor, save the used chip area and lower complexity of place and route.(4) A broad survey of image texture computation characteristic and partial reconfiguration methodological problems and apply the difference-based partial dynamic reconfiguration method to realize image classification.The texture which based on GLCM is widely used in all kind of images'classification and feature extraction. The heavy computation load and long time problem when using software to calculate the texture can only be endured by dimensionality reduction or deflation traditionally. But that obviously lower the amount of effect of follow-up feature extraction. After analysis the proposed image classification algorithm directed towards the characteristics of reconfigurable platform, the system design is completed by using difference-based partial dynamic reconfiguration method which made the calculation of texture and even image classification become quasi real time. System's design document has good portability and possible to realize the other source image classification quickly if a very few modification be made. Results show that when a difference-based partial dynamic reconfiguration method introduced, the system can improve the classification speed and save the used chip area.