Object Detection Based Image Retrieval Algorithm

Along with the development of Internet,Social Networking Service,and We Media,the number of images have an explosive increase and large-scale image databases continue to emerge,which lead to a rapid increase in the demanding of image retrieval.Most of the current image retrieval techniques adopt text-based,content-based,and the combination of these two mixed image retrieval technologies,which are difficult to ful y meet the requirement of users because of the low accuracy and poor efficiency in the image retrieval field.Based on the understanding and analysis of content-based image retrieval technology,a new image retrieval technology is proposed in this paper.This method retrieves images based on the content of object detection,which owns high accuracy and good efficiency.The main contributions of this paper are discussed as follows.1.Several key techniques of CBIR(Content Based Image Retrieval)are deeply analyzed,including the feature descriptions of SIFT(Scale-invariant feature transform),MSER(Maximally Stable Extrernal Regions),and HOG(Histogram of Oriented Gradient).This paper selects SIFT and MSER as the image feature extraction.The similarity measures between the features and the evaluation methods of image retrieval algorithms are described.2.The history of object detection technology is introduced,with the discussion and analysis of several common methods of object detection.YOLO(You Only Look Once)object detection algorithm is applied in this paper,then the basic theory,configuration,training,and usage of YOLO algorithm are introduced.3.In the large-scale image retrieval,we need to match a single SIFT feature with mil ions or billions of SIFT features.In this way,the discriminative power of SIFT feature decreases rapidly,because of many false positive matches between individual features.MSER detects affine-covariant stable regions.Usual y the MSER detector outputs a relatively small regions and their repeatability and distinctness are higher than those of the SIFT.In this paper,we combine SIFT and MSER into a bundled feature.A larger feature has less repeatable and higher discriminatory.On the basis of object detection,image features are extracted to generate bundled features.4.In the large-scale image retrieval,extracting too many features may lead to the curse of dimensionality.This paper selects KMeans algorithm for feature clustering.KMeans algorithm is fast,simple,and has a high efficiency and scalable for large data set.However,there is convergence to the local minimum problem.Therefore,in the quantization process,this paper adopts the k-d tree algorithm to quantify,combining KMeans and k-d tree together to generate the bag of visual word(Bo W),improves the efficiency and accuracy of image retrieval.5.Image inverted index is used in image retrieval.The image features of image database are assigned to the nearest visual word,creating index for the image features to achieve the target of improving the efficiency of the retrieval.The similarity between the query image and the image feature database is measured,then the results are sorted and outputted.The method,which has been proved to be a good image retrieval method by the experime nt results,has advantages of fast speed and high accuracy.