HCNM: Hierarchical cognitive neural model for small-sample image classification

Abstract

Small-sample image classification is a hot topic in computer vision. Despite the progress made by some deep neural networks in solving the small-sample learning problem, there remain challenges in learning efficiently and robustly. These challenges can affect the overall performance and effectiveness of the model. To address these issues, we propose a hierarchical cognitive neural model (HCNM) based on the simulation of visual cognition to construct the sparse structure of the neural model from the perspective of semi-supervised learning. We use a deep learning network for feature extraction and two coupled dynamic neural field equations to simulate the encoding and classification functions in visual image recognition and classification. The model simulates macroscopic neural activation in object recognition and identifies representative point neurons (RPNs) by evaluating the magnitude of lateral interactions within the V4 neural field on an adaptive cognitive scale. Our approach provides an efficient small-sample image classification algorithm that does not require complex parameter tuning and maintains biological plausibility and interpretability. Experimental results using four real-world image datasets demonstrate the superior performance of our model and method for small-sample image classification compared to other state-of-the-art research methods.