孫旭菲 繆新穎 畢甜甜 王水濤 喻芳宇



摘要: 針對因光照、 拍攝角度及圖片質量等因素導致的經典深度學習算法難以有效提取人臉特征、 人物身份識別準確率難以達到理想精度的問題, 提出一種基于人臉強語義的年齡識別算法. 首先, 通過注意力矩陣增強人臉區域的特征權重, 達到提取特征區域的目的; 其次, 使用級聯雙向長短期記憶(Bi-LSTM)網絡學習時序幀之間的特征依賴關系, 彌補部分特征缺失對識別精度的影響. 在人臉數據集IMDB-WIKI和數據集Adience上進行測試, 該算法的年齡識別準確率分別達到78.34%和77.89%. 實驗結果表明, 相比于其他基于深度學習算法的方法, 該算法在基于圖片數據集的人物年齡識別任務上具有更高的準確率.
關鍵詞: 年齡識別; 人臉識別; 深度學習算法; 注意力矩陣; 級聯Bi-LSTM
中圖分類號: TP391.41文獻標志碼: A文章編號: 1671-5489(2024)02-0347-10
SFSR-Age: An Age Recognition AlgorithmBased on Strong Facial Semantics
SUN Xufei1,2, MIAO Xinying1,2, BI Tiantian1,2, WANG Shuitao1,2, YU Fangyu1
(1. College of Information Engineering, Dalian Ocean University, Dalian 116023, Liaoning Province, China;
2. Key Laboratory of Liaoning Provincial Marine? Information Technology, Dalian 116023, Liaoning Province, China)
Abstract: Aiming at the problems that the classical deep learning algorithm was difficult to extract facial features effectively and the accuracy of character identification was difficult to reach the ideal accuracy due to factors such as illumination, shooting angle and image quality, we proposed an? age recognition algorithm based on strong facial semantics. Firstly, the feature weights of facial regions were enhanced by the attention matrix to achieve the purpose of extracting feature regions. Secondly, a cascaded bi-directional long short-term memory (Bi-LSTM) network was used to learn the feature dependency relationships between temporal frames and? compensate for the influence of missing features on recognition accuracy. When tested on IMDB-WIKI facial dataset and Adience dataset, the age recognition accuracy of the algorithm reached 78.34% and 77.89%, respectively. Experimental results show that compared with other methods based on deep learning algorithms, the proposed algorithm has higher accuracy in the task of person age recognition based on image datasets.
Keywords: age recognition; facial recognition; deep learning algorithm; attention matrix; cascaded Bi-LSTM
基于人臉的人物年齡識別算法在未成年人上網防沉迷、 網絡身份識別等領域應用廣泛, 目前, 基于人臉年齡識別方法主要分為兩類: 基于圖像分析的方法和基于回歸分析的方法. 陳文兵等[1]提出了基于多模型融合的年齡識別模型CNN-SE-ELM, 該模型首先使用卷積神經網絡(CNN)初步提取人臉特征, 然后使用SENet(squeeze-and-excitation networks)進一步提取人臉深層特征, 最后使用EM-ELM(error minimized extreme learning machine)算法實現對人臉年齡的分類及性別分類. 實驗結果表明, 該算法能提取更有效的面部特征, 且分類器能快速收斂. 陳濟楠等[2]提出了改進的深度神經網絡模型, 使用大卷積核以及跨連卷積實……