A Research on the Application of Face Recognition Algorithm Based on Convolutional Model and Transformer Model in Community Monitoring System

서지정보

· 발행기관 : 한국사물인터넷학회
· 수록지 정보 : 사물인터넷융복합논문지 / 9권 / 5호 / 61 ~ 72페이지
· 저자명 : 담하의, 민병원

초록

The promotion of intelligent security community construction has greatly enhanced the intelligence and safety of residential areas. In order to further establish a security-oriented community, this paper proposes the utilization of facial recognition based on community surveillance footage to identify suspicious individuals. To address the difficulties in capturing facial images caused by factors such as low pixel resolution and varying shooting angles in surveillance footage, the following optimization strategies are proposed in this paper : Firstly, a lightweight global search facial detection network is designed based on convolutional modules and Vision Transformer modules. The Vision Transformer module is introduced to enhance the global retrieval capability of the network. Secondly, the structure of the Vision Transformer module is optimized by adding pooling layers in the feature block extraction and segmentation stage to reduce the number of module parameters. The feature blocks are mapped and computed with the feature maps to improve the corresponding feature correlation. Thirdly, in the face alignment stage, an Anchor Free mechanism is adopted to generate elliptical face localization regions for more accurate fitting of faces and reducing interference from other background information in the final identity recognition stage. Finally, the similarity between faces is calculated using Euclidean space distance to determine corresponding personnel identities. Through relevant experiments and tests on the self-built facial identity dataset in this paper's residential surveillance system, the proposed facial detection network achieves an average improvement of 3.11% in detection accuracy compared to other detection networks, reaching 97.19%. In terms of facial identity recognition, the designed model achieves an average improvement of 3.43% with a recognition accuracy of 95.84.

참고자료

· 없음