Extraction And Recognition Of Objects In X-ray Imaging

X-ray imaging technology has been widely used in medical,industrial production and manufacturing,public safety,food safety and other fields,the X-ray image processing has become a research hotspot with the development of computer technology and hardware equipment.In this dissertation,two kinds of X-ray imaging images,cardiac angiography image and security instrument image,are processed.The sparse representation,dictionary learning and bag of visual words technology are utilized to extract and recognize the object in those images.The main contributions of this dissertation are summarized as follows:(1)A new algorithm based on multi-dictionary and sparse representation is proposed to solve the problem of poor contrast in small blood vessels and hard for recognition in cardiac angiography images.The proposed algorithm uses the training dataset to get the Representation Dictionary(RD)firstly,and the Enhancement Dictionary(ED)is generated by manual extracting results from the corresponding blood vessel regions in the training dataset.The RD and ED are optimized according to the proportion of the vascular region in the single atom.Then,the image in the test dataset is represented by RD,and the sparse solution of each small patch is solved via the Orthogonal Matching Pursuit(OMP)algorithm.The enhanced image is constructed from ED and the solved sparse solution,and a gray-scale stretching method is utilized to gain the final result.The algorithm can not only improve the contrast ratio of the target region in blood vessel image,but also can improve the structure of small blood vessels and enhance the details.In the experimental part,we also apply the proposed algorithm to the retinal vessel image enhancement.(2)In order to improve the precision of blood vessel area extraction,a new algorithm of blood vessel extraction based on sparse representation and dictionary learning is proposed.When the training data set is used to generate the segmentation dictionary,all the vascular regions in the training data set are extracted to obtain a mask,and the segmentation dictionary sets are generated according to the mask information.Then,we add the label information and the center position of the dictionary to each atom in the obtained segmentation dictionary sets,and a dictionary learning method is also employed in the proposed method.When the test image is being processed,several segmentation dictionaries are selected from the segmentation dictionary sets according to the location information,then the test image is represented by these selected segmentation dictionaries.The solved sparse solution and label information are used to determine whether the small blood vessel patch is the blood vessel region,then the result is obtained.The experimental results show that the proposed method can extract the blood vessel region with high precision.The same as vessel image enhancement,we also apply the proposed algorithm to the retinal vessel image segmentation.(3)For another type of X-ray imaging image,security instrument image,we achieve the object region extraction of the control tool and dangerous goods in the security instrument image.A non-local active contour model used for object extraction is proposed,which uses a non-local method to compute the Gabor feature map and level set method to extract the object region.The experimental results show that the structure of the object region obtained by the proposed method is complete,the contour information is similar to the ground truth,and the algorithm shows less sensitive to the initial contours.The algorithm is also applied to the extraction of the tumour object region in a liver CT image.(4)A security instrument image recognition method based on bag of visual words and sparse representation is proposed to solve the problem of low accuracy on security instrument image recognition.The proposed algorithm is divided into recognition dictionary generating part and test image recognition part The visual words are extracted via Speeded Up Robust Features(SURF)operator.The extracted visual words are filtered according to the ratio in the training dataset,and the common characteristics of different objects and the impact of noise are removed through that filtering.Then,the bag of visual words are generated from k-means cascade clustering method,and a recognition dictionary with label information is obtained from the generated bag of visual words.In test image recognition part,we use the SURF operator to obtain the visual word,and use the recognition dictionary to perform the sparse representation.Then,the sparse solution and the dictionary label information are utilized to obtain each visual word recognition result,and the final recognition result of the test image is obtained by an operation of voting.The experimental results show that the proposed algorithm has high accuracy and fast computation speed,which can achieve the security instrument image recognition.