Image Retrieval Based On Feature Performance Enhancement And Target Localization

With the rapid development of Internet technology and the popularity of digital products such as mobile phones and cameras,the number of images that circulating on the Internet is increasing every second,and the content of images is becoming more and more colorful.Quickly and accurately retrieving the images,whitch we need,in large image database has become a hot research topic.How to extract image features with strong characterization performance and how to accurately locate the region of interest is the focus of research in the image retrieval field in recent years.Based on this background,this paper proposes a new method,whith is image retrieval based on feature performance enhancement and target localization.There are two main innovations here:The first innovation is that we propose a performance enhancement method for convolution features by improving the CroW feature weighting method.After channel weighting and spatial weighting,we obtain multiple local feature vectors by sliding on the image convolution feature maps with multi-scale pooling windows.Then integrate this vectors into global feature vectors.Experiments show that this feature vectors have strong image representation.The second innovation is that the paper firstly combines the object detection method,which based on deep learning,with the image retrieval method,and proposes an image retrieval method based on target location.We use the retrained SSD object detection model to locate the desired target region in the image,and extract the region features according to the target regions for image retrieval.Because the method extract features on specific regions of the image,the image interference factor can be filtered out to the greatest extent,and the retrieval precision is greatly improved.The main contributions of this paper are three.The first contribution is that we obtain a target detector that accurately detects the building regions of images by useing the hand-labeled building dataset to train the SSD model.The second contribution is that the convolution feature performance enhancement method can greatly improve the retrievalThe second contribution is that the image retrieval method Based on the target location have a mAP value(average mean mean value)of 90.3%and 80.1%on the Paris6k database and the Oxford5k database,respectively.In terms of retrieval accuracy,this method is superior to most image retrieval methods in recent years.