Specific Person Recognition Based on Local Segmentation and Fusion

With the development of video and image technology, it is of great practical value to recognize specific persons in television programs and photo albums. However, occlusion of body parts and changes in shooting position and distance are common in real scenes. In this work, we proposed a Specific Person Recognition based on Local Segmentation and Fusion method, called PR-LSF, which improved the reliability of person recognition in these environments. We represented the human body as an aggregate of multiple parts and apply local segmentation to train multiple convolutional neural network (CNN) classifiers. Each part classifier generated an identification decision confidence for each part. By training the SVM classifier, we weighted the decision confidence of all parts to make a comprehensive judgment. To verify the effectiveness of the proposed algorithm, we performed experiments with unoccluded and occluded test sets. The experimental results demonstrated that PR-LSF achieved higher recognition performance than algorithms using a single body part and were reliable even with partial occlusions, multiple scenes, and shooting changes.