While Convolutional Neural Networks (CNNs) excel at image recognition, they require extensive training data, in contrast to humans, who can learn from just a few examples. This study proposes a method that simulates human visual processing to enable few-shot learning. It generates "object concepts" by quantifying shape using P-type Fourier descriptors and color using average RGB values. A CNN is then trained using these object concepts as labels, effectively serving as a specialized encoder for low-level features such as shape and color. For classification, a fuzzy inference system functions as a robust decoder. This approach aims to reduce the amount of required training data while maintaining high recognition accuracy. We validated our method on a public dataset containing 222 objects across 37 categories, confirming its effectiveness compared with a conventionally trained CNN.
Shintani et al. (Mon,) studied this question.