AniWho : A Quick and Accurate Way to Classify Anime Character Faces in Images
This work addresses a domain-specific problem for anime image classification, but it is incremental as it applies existing methods to a new dataset.
The paper tackled the problem of classifying anime character faces by evaluating models like EfficientNet-B7, MobileNetV2, and Prototypical Networks, with EfficientNet-B7 achieving the highest top-1 accuracy of 85.08%.
In order to classify Japanese animation-style character faces, this paper attempts to delve further into the many models currently available, including InceptionV3, InceptionResNetV2, MobileNetV2, and EfficientNet, employing transfer learning. This paper demonstrates that EfficientNet-B7, which achieves a top-1 accuracy of 85.08%, has the highest accuracy rate. MobileNetV2, which achieves a less accurate result with a top-1 accuracy of 81.92%, benefits from a significantly faster inference time and fewer required parameters. However, from the experiment, MobileNet-V2 is prone to overfitting; EfficienNet-B0 fixed the overfitting issue but with a cost of a little slower in inference time than MobileNet-V2 but a little more accurate result, top-1 accuracy of 83.46%. This paper also uses a few-shot learning architecture called Prototypical Networks, which offers an adequate substitute for conventional transfer learning techniques.