Research On Large-scale Face Image Retrieval On Deep Learning

In today'world,the rapid development of Internet technology affects every aspect of people's life.In this era of frequent information exchange,a large amount of image data is accumulated in the Internet.How to effectively use and organize these image data has become an urgent problem to be solved.This has also spawned the rise and development of content-based image recognition and retrieval technologies.As the number of images increases,image retrieval accuracy and retrieval speed face enormous challenges.The hash method is favored in image retrieval technology for its high efficiency and low memory usage.At the same time,thanks to the rapid development of deep learning technology in the field of computer vision,deep hash technology has gradually become the mainstream method in the field of image retrieval.In the known end-to-end depth hashing technique,one only pays attention to the similarity between images,and ignores the semantic information of the image itself.Moreover,in the large-scale data training,the influence of the label space on the training process is neglected,which makes the model difficult to converge during the training processIn view of the above problems,this paper improves the current mainstream deep hashing method,and proposes a deep hashing method that combines the semantic information of the image itself with the similarity of the image to improve the performance of retrieval models.And using asynchronous training method and optimized data generation solves the problem that the network does not converge and the training is unstable during the training process of large-scale data and large-label spatial data,and the retrieval precision is improved.This paper further designs and implements a multi-process multi-level parallel retrieval scheme based on GPU,which speeds up the retrieval speed on large-scale data sets.The specific work of this paper is as follows1)A multi-task deep hash method is proposed.On the existing deep hash method,a deep learning method combining classification task and hash coding task is adopted to combine classification loss and hash loss.The semantic information of the picture is used to guide the learning of the hash coding task.Finally,an end-to-end feature extraction and hash coded multitasking network is obtained.Use the MS5W dataset(the proprietary data set of Zhongke Shi Tuo Company,which contains more than 50,000 face labels and 3.4 million face image data)to train,The dataset of MSAS(The dataset is a hybrid of the Microsoft Face Dataset and the Green Celebrity Asia's public celebrity face dataset.It has more than 90,000 face tags and 5.1 million face image data)shows that the method is effective.The model performance has been improved by 2 MAP(Mean Average Precision)indicators2)The network structure with better performance is designed and implemented.Based on the ResNet network structure,a more reasonable feature coding method is used to make the training process of the network more stabnetworkle and the convergence speed is faster.Higher retrieval accuracy can be obtained using this.The network tested and improved 1-2 MAP metrics on the MSAS data set3)Using Triplet Loss instead of Softmax Loss as the classification loss function solves the problem of difficult training of large-label spatial data.The design and use of an efficient data input method makes the training of the network more stable,using asynchronous training and online generation of triples,so that the network can converge quickly4)Based on Hamming's sorting,a multi-threaded parallel retrieval strategy is proposed,which makes the retrieval speed faster and the precision higher.This makes it possible to implement fast and efficient retrieval methods with limited computing resources.In the end,the single image retrieval speed in the million-level database reached 1.4ms.