Research On Deep Learning-based Face Attribute Recognition Algorithm

In recent years,the field of computer vision has seen rapid development,with facial attribute recognition technology being one of the research hotspots that has been widely applied in various areas of life,such as video surveillance,intelligent security systems,and commercial internet advertising recommendation models.At the same time,with the emergence of massive facial image data and the increasing demand for image analysis tasks,the application scenarios of facial attribute recognition tasks have also been constantly enriched and expanded.These trends provide a wide range of application scenarios for the field of computer vision and provide ample space for the development of artificial intelligence technology,Against this backdrop,leveraging the advantages of artificial intelligence to further explore facial features,improve facial attribute recognition performance,and address fairness issues in recognition has extremely important practical significance.Given the unique features of facial attributes,it is worth exploring the combination of global and local features to improve attribute recognition accuracy.Different from existing methods,this paper proposes a novel Transformer-based facial age recognition network that includes a Feature Mining Module.The Feature Mining Module can adaptively extract global features and local features at different scales,and perform various combination-based fusion operations on the extracted features to enhance the information interaction between blocks,Meanwhile,the network adopts a strategy of mutual learning within the model to enhance the interaction between global and local information,further improving the model’s recognition accuracy.The good results achieved on three public datasets have verified the effectiveness of this model.In addition,this paper has constructed a dataset named Labeled Gender in-the-wild(LGW)in a real-world environment to address issues such as skewed data distribution and insufficient diversity in the current dataset.The LGW dataset contains one million facial images of 200,000 subjects from around the world,and is unbiased in terms of skin color and gender attributes.The existence of the LGW dataset provides an experimental foundation for further research on fairness in face attribute analysis.In practical applications,multiple attribute information in a face is interdependent and difficult to fully decouple.As a result,the target task is susceptible to bias predictions influenced by other factors.Therefore,this paper introduces a fair-aware algorithm for facial attribute recognition,which designs a Sensitive Attention Module and a Mutual Learning strategy for the target attribute.The algorithm reduces attention to sensitive information by identifying high-response areas of the Sensitive Attention Module positioning-sensitive features,and the strategy obtains more stable and fair feature representations by learning the invariance of unbiased features.Experiments on LGW and three public datasets have verified that the algorithm can effectively eliminate bias and improve recognition accuracy.