Multimodal Camera-Based Gender Recognition Using Human-Body Image With Two-Step Reconstruction Network

With the recent development of intelligent surveillance systems, the importance of research study on gender recognition of people at a distance is also on the rise. The existing gender recognition technologies studies have used high-resolution facial images captured from the front at a short distance, which showed high performance. However, intelligent surveillance systems in actual real environments have difficulty in detecting the faces of people because they use images captured from a distance. Moreover, in the case of back-view images, gender recognition based on the facial image is impossible because the face cannot be detected. Thus, gender recognition using the full-body human-body images of people is being studied but its performance is low owing to problems such as low resolution, motion blur, and optical blur. Furthermore, the performance of gender recognition using only visible-light cameras is limited owing to illumination variations, shadow, and the type of clothes and accessories. To solve these problems, remote body-shape-based recognition was performed by sequentially using two convolutional neural networks which improved the resolution of visible-light images. In addition, the degradation of recognition performance owing to various factors (e.g., illumination, shadow, and the type of clothes and accessories) was prevented by combining a visible-light camera with an infrared camera, and the scalability was enhanced using various heterogeneous cameras. The higher performance of the proposed method compared with that of other methods was verified through a comparative experiment using the open database of Sun Yat-sen University multiple modality Re-ID (SYSU-MM01) and the Dongguk body-based gender database (DBGender-DB2) that has been built by us.