The Machine Learning Research For Image Retrieval

Over the last few years,with the rapid development of the Internet,individuals can obtain countless multimedia resources.Compared with other multimedia resources,images can express things more directly,and the content is rich without any language restriction.Therefore,massive digital images are produced on the Internet.How to efficiently and accurately query the images in which people are interested in the massive database becomes one of the popular directions in computer vision research.Influenced by the application of CNN in image recognition,detection and classification tasks,some studies have applied CNN to image retrieval in recent years and achieved great success.However,the existing research extracts the fully connected layer output as image features,and then performs statistical-based mathematical methods to measure similarity.The image retrieval model based on these steps makes the image features incomplete and does not make full use of the convolution network.The retrieval takes a lot of time,which leads to unsatisfactory real-time retrieval efficiency.Aiming at the above problems,this paper studies the application of machine learning in image retrieval in detail,and uses the latest deep learning model to realize the architecture of image semantic feature and similarity measure fusion.The image retrieval research in this paper mainly consists of two parts: First,the fusion deep learning and nearest neighbor algorithm based on the nearest neighbor image retrieval system,Secondly,the image retrieval is based on the classification,the classifier is built as feedback information to establish a model suitable for image retrieval.(1)Establish a model of deep learning and nearest neighbor algorithm fusion.Firstly,the transfer learning method is used to fine-tune the Convolutional Neural Network parameters,the nearest neighbor algorithm KD-Tree is used to measure the similarity between images,the query image returns the image by querying and backtracking the KD-Tree root node to the leaf node.Then,we established a deep learning model based on image fine-grained features.The neural network p arameters are adjusted by the truple loss function and applied to the vehicle fine-grained recognition retrieval.In the Open Data V11.0-Vehicle Re-identification Dataset,we established respectively a deep learning single-label based retrieval model,a deep learning multi-label based retrieval model and a deep learning-based fine-grained retrieval model.The experiment proves that by extracting the microscopic features of the vehicle,the image of the same car can be effectively retrieved,and the real-time roa condition vehicle safety monitoring under the condition of license plate recognition failure is effectively realized.(2)We propose a simple yet practical approach,namely deep learning of preclassification(DLPC),which integrates classification into the image retrieval framework.Specifically,DLPC learns a Convolutional Neural Network(CNN)model via transfer learning,which can simultaneously finish feature extraction and image preclassification.The results of image pre-classification provide feedback information indicating that features of images belonging to the same class should be stored together.For new-coming query images,they can easily find their similar images from the image library,as the similarity metric is performed only on images in the same category according to the pre-classification results.Extended experiments are performed on the Wang dataset and the Pet Dataset.Experimental results on these two large scale image datasets validate the promising performance of our method compared with the state-ofthe-art approaches.(3)On the basis of the previous research,we applied the innovative retrieval model of the fusion image classification to the marking system,which reduces the difficulty and cost of manually assigning the questions to the positioning point.It provides a new approach for a more efficient intelligent scoring system.