Instance-Aware Image Retrieval Technology Based On Multi-Task CNN

With the popularization of mobile Internet and digital media,image and video data have gradually become an indispensable resource for the network.The effective storage and efficient retrieval of massive images in the multimedia era are our current opportunities and challenges.In this challenge,image hashing is the key to mapping high-dimensional data into low-dimensional binary hash codes.The quality of the hash code depends on the design of the hash function in the image hash algorithm and the validity of the image representation.Image features based on manual extraction have limited ability to represent images,and the results are not ideal in practical use.The design of hash functions is also difficult to apply to multiple data sets.In recent years,computer vision and deep learning have developed rapidly.The method of automatically learning data features based on deep learning has gradually replaced the original manual acquisition,achieving a smarter and more automated feature extraction method.The combination of deep learning technology and hash coding technology is the mainstream research direction of current image hashing technology.In the depth image hashing technique,it is a research hotspot to find and explore more effective convolutional neural network structures and optimization algorithms.According to the research and analysis of the status quo of deep learning image hashing technology by domestic and foreign scholars,this paper mainly analyzes and experiments the depth image hashing technique of image retrieval.The work is as follows:Aiming at the low level of automation and intelligence in the existing map search technology,lack of artificial intelligence technology support,difficulty in obtaining accurate search results,large storage space consumption,slow speed and difficulty in meeting the image retrieval needs of the big data era.A multi-task cascaded hierarchical image retrieval method is proposed.Firstly,the selective retrieval network is used to perform logistic regression on the feature map to obtain the probability vector of each region of interest in the image.Based on this,the compact quantization network is combined with the compact quantization network to obtain a compact quantization hash code.Secondly,the network is again filtered to obtain each.The regions with the most responsive regions of interest are aware of the semantic features;then,each region of interest is subjected to a precision search strategy based on the quantized hash matrix to quickly compare the images;finally,the most similar to the corresponding regions of interest in the query image is selected.image.The proposed multi-task learning method can obtain the compact quantization hash code and the region-aware semantic feature of the image at the same time and can effectively remove the interference of the image background and other object information.The experimental results show that the proposed method can achieve end-to-end training,and automatically select higher quality regions of interest to improve the intelligence level of large-scale image retrieval.The retrieval accuracy(0.9478)and retrieval speed(0.306s)are obviously better than those.Existing large-scale image retrieval technology.Aiming at the lack of intelligence in existing vehicle retrieval and the fact that the retrieval results can not meet the actual needs,a bayonet vehicle retrieval technology based on multi-tasking network for segmentation compact image extraction of fine-grained data sets is proposed.Image diversity and relevance complete online retrieval of road bayonet.Firstly,the accuracy of the retrieval system is improved by combining the related tasks,and the coupling between multi-label features is reduced.The proposed multi-label segmentation learning strategy combines semantic features and multi-labels to obtain binary hash codes of different attributes of vehicle images.Then,the robustness of image representation is enhanced by minimizing image coding.Then,the example features of the vehicle image of the feature pyramid network map are adopted.Then,the local sensitive hash reordering algorithm is designed to match the similar features of the above features.Finally,A cross-modal approach is used to aid retrieval in the special case of missing or missing retrieval vehicles.Using the retrieval methods proposed in this chapter on three public datasets,the retrieval accuracy is better than the current mainstream vehicle retrieval algorithms.Among them,the search data of CompCars vehicle data set is 0.966.The vehicle ID is raised to 0.862.The method minimizes image coding and image instance feature fusion to obtain high-precision and high-aging retrieval results and can realize cross-modal retrieval in the absence or defect of vehicle images.