Study On The Image Processing And Recognition Of Low-quality Visual Images

There exist large amounts of low-quaility visual images in the real world and they have been used widely. There are many factors which lead to the image blurring, serious detail loss, image warping and ambiguities images, for those reasons, the images we observed are usually not accurate enough to fit the requirement of the image resolution. It is this characteristic that makes it difficult to use the existing research achievements of the traditional optical image recognition to recognize them. Accordingly, recognition of such image becomes a new hot spot. Supported by Doctoral Fund of Ministry of Education of China, the research carried out in this dissertation is as followings.The low-quaility visual image preprocessing is investigated first. Aimed to solve the problem cause by low-resolution, skewed image, rotated image and broken image, the reconstruction of broken characters, filling the holes of image, image enhancement are researched. As the image preprocessing is the precondition of image feature extraction and recognition; such researches paly an important role of theoretical guidance.A novel low-quaility visual image recognition based on the small region template-matching algorithm is presented. It is used to overcome the shortcoming of long time waste and unsuitable for the industrial process. The laser etched label characters and the deep space objects are used to testify the probability of the method, it shows that this algorithm uses less time and could get much higher recognition rate. Based on this method, the theory of correlation factor is lead into the small region template matching method to solve the problem of size mismatching. The experiment shows that this method is suitable for the motion objects; it also shows that this method can easily overcome the deformation, noise and local shading, so this method has much higher matching precision.The rotated image will decrease the recognition rate by using principal component analysis. To solve this problem, a novel sample vector formation method is presented. The invariant features such as moment invariants are used to get the rotate invariant vectors; the recognition is carried out directly on the gray-level images by adopting the improved PCA subspace method. Experimental results show that this method could decrease the number of sieving samples and has much higher recognition rate comparing with the typical method.A central issue in PCA is choosing the number of intrinsic components to retain. However, most studies assume a known dimension or determine it heuristically, though there are a number of model selection criteria in the literature of statistics. In this dissertation, the probabilistic reformulation of PCA is used and a model selection criterion for determining the intrinsic dimensionality of data including Akaike's information criterion (AIC), the consistent Akaika's information criterion (CAIC), and the Bayesian inference criterion (BIC) are derived. At last, aimed to the character images, these parameters could affect the determination of the subspace dimension is analyzed in detail. The range of application of these three criterions is analyzed and a two-step method to estimate the intrinsic dimension of the observed character dataset is presented. Experiments demonstrate that the new algorithm is feasible and robust; it also can decrease the time waste. The comparison experiment for the recognition of protuberant characters shows that this method has much higher recognition rate than typical used PCA methods.A new algorithm based on the structural properties such as the endings, the three crossing points, the four crossing points and the positions of the three crossing points is presented. It could solve the problem of low recognition rate caused by rotate, similar and broken characters. A two-level recognition network is setup, the first network is based on the PCA method, after the first step recognition is complete, the result is sent to the second network which is based on the described structural features, the error-prone characters and the character with low degree of confidence are recognized again. Experiment results show that the two-level character recognition network has much higher recognition rate and it is much robust as it can recognize the similar characters and some broken characters.A single sensor cannot produce a complete representation of a scene, Image fusion is the process of combining information from two or more images of a scene into a single composite image that is more informative and is more suitable for visual perception or computer processing. A novel fusion method of low-quality images based on hybrid Contourlet-PCA transform is proposed. The experiment result shows that the proposed method could well fuse hyperspectral images with noises eliminated and it outperforms the Contourlet and PCA methods.This work is supported by Doctoral Fund of Ministry of Education of China.