Animation character recognition based on weighted improved single shot multi-box detector model

Abstract

Aiming at the low accuracy and poor adaptability of animation character recognition in complex scenes, an intelligent recognition method based on weighted improved single-detection multi-frame detector (SSD) is proposed. By constructing a special data set containing 36 types of animated characters, the positive sample dynamic weighting strategy is innovatively designed to solve the inherent positive and negative sample imbalance problem of SSD model, and the receptive field enhancement module (RFB) is introduced to improve the multi-scale feature expression ability. The experiment shows that the mAP of the model increases by 13.8%. When the positive sample weight coefficient is 3, the identification effect is the best, which is 18.2% higher than that of traditional SSD. The improved SSD-RFB model detection accuracy (mAP@50 = 95.78%) and real-time (10.5 FPS) are superior to mainstream detection algorithms. The results show that the proposed algorithm can effectively alleviate the overfitting problem caused by the pose diversity of animation characters, and the synergistic effect of core modules is demonstrated through experiments. The research provides an efficient solution for the intelligent retrieval of animation resources, and its sample balance strategy and feature enhancement method provide a new way of thinking in the field of animation character recognition.