Research on the Application of Improved SSD Algorithm in the Intelligent Recognition Model of Anime Characters

In order to apply object detection techniques with better detection performance to anime character detection, an improved Single Shot MulmtiBox Detector (SSD) algorithm is proposed. Based on the basic SSD algorithm, the research introduces a Feature Pyramid Network (FPN) upsampling module in feature layer prediction, and replaces the VGGNet network structure in the basic algorithm with ResNet network structure as the front-end network structure. And introduce weight parameters to adjust the proportion balance in the samples, optimize the training strategy of the algorithm, and obtain the final improved SSD algorithm, further improving the detection accuracy and speed of the algorithm. The study compared the performance of the improved SSD algorithm with other algorithms on the COCO dataset and the anime character dataset. The results showed that the improved SSD algorithm achieved an average accuracy (mAP) of 0.78, a positioning error of 0.04, a recall rate of 0.87, and a processing speed of 18 FPS on the COCO dataset, all of which were the best among the algorithms. The algorithm achieved a detection AP value of 0.82 for character 30. The average detection time of the improved SSD algorithm is 32ms, and the MAP reaches 0.74, which is the highest among all algorithms. At the same time, experiments were conducted on the value of the weight parameter. When the weight parameter value was 1, the detection mAP value remained stable at 0.77, achieving the best detection effect. The results show that the improved SSD algorithm proposed in the study has superior detection performance, and has further improved detection accuracy and speed, providing a certain experimental basis for the detection research of anime characters.