Research And System Implementation Of Multi-object Hashing Method For Social Media Images Based On Weak Supervision

Content-based image retrieval is one of the important research technologies in the field of information retrieval.It can retrieve the required data from a large amount of image data and plays an important role in the context of the era of big data.One of important application scenarios is social media image retrieval.In the long-term development process,mainstream methods use a single feature to represent images,and its inherent defects lead to An unilateral and poor outcome.There are still barriers when apply retrieval in actual life,the poor training due to lack of artificial tags and slow retrieval response due to dimensional disasters will also occur at the same time.The present study attempts to solve the many problems in the actual application of social media image retrieval,a method to solve the overall problem is proposed,and a method of image multi-object hash retrieval based on weakly-supervised detection is realized,and the image is learned by using social weak tags.The hash representation of each target in the solution solves the problem of supervised learning's dependence on artificial tags,and can also effectively improve the retrieval effect and enrich the retrieval form.The main work content is as follows:(1)In order to effectively use the information in social tags,research has been carried out on the problems of noise,lack,and expression differences in social tags,and corresponding optimization solutions have been proposed.The Word Net model is introduced to deal with noise according to semantics.In processing,the social tag completion optimization is performed by fusing the correlation between social tags and the similarity between image visual features,and the problem of expression differences is solved through semantic clustering,and the optimization of social tags is completed from many aspects.Improving the quality of social labeling.This process will improve the quality of social labeling.(2)Nowadays,the social media image retrieval methods mainly based on the single feature encoding representation of the image which leads to poor performance and unilateral retrieval outcome especially in multi-target image retrieval tasks.An image multi-target hash model based on weakly supervised detection is proposed.By constructing a multi-task deep learning network architecture,using optimized social tags,the image target area detection and target hash representation are learned in two branches.The characteristics of the image are represented by a target hash set,which can take into account the requirements of image retrieval for time and accuracy and can effectively expand the form of image retrieval.(3)In the two task branches of weakly supervised detection learning and hash learning,they will face to label inaccuracy problem.For this,a label-weighted loss function is designed to reduce the interference of noisy labels on model training,and the two related tasks are combined with a common optimization loss function to complete the training of the multi-task model.A large number of experimental comparison results prove that the label optimization method in this paper can significantly improve the quality of social labels,and provide good basic support for the model training of retrieval tasks.Based on the proposed image multitarget hash model based on weakly supervised detection,the hash feature representation of each target in the image can be effectively obtained.Under multiple image retrieval evaluation indicators,it has obvious advantages compared with the current mainstream social media image retrieval methods,and compared with the actual retrieval results,the retrieval results are more in line with actual retrieval requirements.In the end,relying on the image multi-objective hash retrieval method proposed in this article,combined with the actual image retrieval requirements,designed and implemented a fully functional social media image retrieval system,which has a better user experience.