Pedestrian Gender Recognition by Style Transfer of Visible-Light Image to Infrared-Light Image Based on an Attention-Guided Generative Adversarial Network

Abstract

Gender recognition of pedestrians in uncontrolled outdoor environments, such as intelligent surveillance scenarios, involves various problems in terms of performance degradation. Most previous studies on gender recognition examined recognition methods involving faces, full body images, or gaits. However, the recognition performance is degraded in uncontrolled outdoor environments due to various factors, including motion and optical blur, low image resolution, occlusion, pose variation, and changes in lighting. In previous studies, a visible-light image in which image restoration was performed and infrared-light (IR) image, which is robust to the type of clothes, accessories, and lighting changes, were combined to improve recognition performance. However, a near-IR (NIR) image requires a separate NIR camera and NIR illuminator, because of which challenges are faced in providing uniform illumination to the object depending on the distance to the object. A thermal camera, which is also called far-IR (FIR), is not widely used in a surveillance camera environment because of expensive equipment. Therefore, this study proposes an attention-guided GAN for synthesizing infrared image (SI-AGAN) for style transfer of visible-light image to IR image. Gender recognition performance was improved by using only a visible-light camera without an additional IR camera by combining the synthesized IR image obtained by the proposed method with the visible-light image. In the experiments conducted using open databases—RegDB database and SYSU-MM01 database—the equal error rate (EER) of gender recognition of the proposed method in each database was 9.05 and 12.95%, which is higher than that of state-of-the-art methods.