Research On Related Technologies Of Computer Vision-based Object Image Retrieval

Computer vision-based object image retrieval is one of the most challenging researchtopic in the field of computer vision, Its main content is how to use computer simulation ofhuman vision to describe the content of the image, and identifying a target image of interestfrom the mass of the image according to the features of the content description. It has beenwidely used in network image search, medical image mining, content-based video retrieval,security monitoring, and pornography image filtering, and is hot and difficult interdisciplinaryresearch in machine learning, pattern recognition, computer vision and image processing. Dueto its complexity and structural, the object image retrieval based on computer visiontechnology still has many challenges, and its accuracy needs to be improved.This paper conducts a series of research on computer vision-based image retrieval, andcombines multi-scale space theory, object significance theory and vision conformity theory topropose feasible scheme for the actual image retrieval application, such as the object imageretrieval of illumination changes, the object image retrieval based on visual consistency andso on. This paper is focused on the feature extraction of the retrieval process, retrieval modeland visually significant problem, and proposes the corresponding solution. Its main researchcontent and innovation are as follows:In feature extraction process, due to the acquisition of conditions change and noiseinterference, the angle variation of object image occurs and illumination changes, theextracted texture features will change, and the image retrieval method based on LBP (LocalBinary Pattern) texture characteristics will probably lapse. To solve this problem, this paperproposes an object image retrieval method based on LBP local texture features of color space.LBP features of color space are actually the combination of simple LBP feature and colorspace feature. It retains the original LBP local feature rotation invariance, and increase theillumination invariant too. In addition, the original LBP features discarding color informationis to reduce the dimension of the feature to improve the retrieval speed. This paper proposes acolor space mapping method which can reduce the feature dimension, and then reduces thecomputational complexity. The experimental results show that object images of complexillumination changes, angle transform and scale changes can be more effectively used forclassification by this method.In the process of feature extraction and image retrieval, a marked out image of objectregion can be helped to improve the accuracy of the target image retrieval. The traditionalmethod is through artificial division of object to train the classifier, but for large capacity image database, such artificial labeling method is not feasible. To solve this problem, wepropose an object image retrieval method based on multi-scale visual saliency. This method issuitable for unsupervised object image retrieval method with simple background and noshelter. At first, we use statistical learning to train a multi-scale target saliency detectionmodel, and then use the model to adaptively extract the significant region in the trainingimage, so we extract the color, intensity and direction of significant area.Finally, we combinethese feature with Probabilistic Latent Semantic Analysis (PLSA) model for the object imageretrieval. The experimental results show that the proposed method can automatically detectthe significant target region in the image, and be effective to improve the accuracy of imageretrieval system by the object significant feature sorting.The results returned from image retrieval system often appear with the unrelated images,and that is not satisfactory. This is because when the significance of local noise images arehigher than the true images, the noise images will become higher ones of image retrievalresults in sort-priority order.To solve this problem, we propose an image retrieval methodscombined with vision conformity and object significance by mining the vision conformityfrom results of image retrieval system. Firstly, we calculate the object saliency maps fromthe results of image retrieval system, and use objective significant coefficients for the initialfilter, and them we strike a vision conformity mode with all the significant object in thecollection images. Finally, we make image clustering according to vision conformity.Experimental results show that this method can not only improve the accuracy of imagesearch results, but also effectively enhance performance of image retrieval system, and givepriority to these images which have visual significance and are closely related to the querytopics image back to the user.