Research And Implementation Of Visual Privacy Object Detection Based On Few-Shot Learning

With the development of intelligent technology,intelligent devices bring convenience to people’s economic and social life.There is a great risk of information leakage and illegal use in the dissemination and use of massive data.Frequent personal information security issues may bring huge property losses and even personal injuries to users.At present,based on more mature biometric protection technologies such as face and fingerprint,people are increasingly concerned about the potential multimedia information security and privacy protection issues in intelligent device applications.The premise of privacy protection is the accurate detection and localization of private content.Among them,visual privacy has characteristics of privacy and non-publicability.The collection and annotation of sample are faced with great difficulties.At the same time,the definition of privacy content is affected by individual subjectivity and scene variability.However,traditional deep learning-based models are limited to the recognition of fixed categories,which cannot meet the needs of sample updating and model generalization.Therefore,how to detect and locate the private content with a small amount of labeled data has become a critical problem.In view of the problems in few-shot learning and visual privacy object detection,the main contribution of this thesis are as follows:(1)A few-shot visual privacy object detection algorithm based on deep aggregation pyramid network is proposed.This thesis fuses classical object detection and few-shot learning methods,designs network sharing methods to retain the basic parameters.With the strong knowledge representation capability of existing deep learning methods,a deep aggregation pyramid structure is designed as the backbone network.Through the deep fusion of semantic and spatial feature,more representation information and global features are extracted,enabling the model to quickly learn richer useful knowledge from a small number of samples.Therefore,this method can solve the problems of few privacy samples,costly labeling,and novel class recognition.The effectiveness of the proposed algorithm for few-shot object detection on different datasets is verified to achieve accurate localization and recognition of visual privacy objects.(2)A few-shot visual privacy object detection algorithm based on cosine compensation loss is proposed.The accuracy of the model is improved by optimizing the classification loss function while ensuring the objects are accurately localized.This thesis integrates the ideas of normalization processing and margin optimization,introduces a boundary strategy based on the cosine cross-entropy function and proposes a weight-based compensation constraint term.This innovation makes the maximum intraclass variation less than the minimum interclass variation and further solves the sample imbalance problem.The propose of this algorithm is to guide the model to learn efficiently in few-shot labeled data to improve the model detection accuracy.And the experiments show that the proposed method achieves some improvement on both PASCAL VOC datasets and COCO datasets.(3)A multi-task based visual privacy object detection and protection system is proposed.Since the detection of inappropriate visual privacy content in practical applications is only focused on simple binary classification,this thesis proposes a joint training strategy based on deep learning.Homemade inappropriate visual content dataset and retain the multi-label output form.The detection of inappropriate visual privacy such as large area of body nudity is achieved while ensuring the effectiveness of base class detection.And the generalization ability of the algorithm is further improved so as to meet the practical application requirements.Finally,for different definitions of visual privacy content,we implement a visual privacy content detection and protection system in specific scenarios.The system is divided into face object detection,inappropriate visual privacy object detection and custom visual privacy object detection,and the recognized privacy objects are masked or noisy to achieve real-time privacy content protection.The privacy content is protected while meeting the overall data availability.