A Fine-Grained Recognition Model Based On Hippocampal Neural Circuit

Fine-grained image recognition tasks have received much attention in recent years. Works in this fields generally suffers from the following drawbacks: firstly, models are designed to be overly bloated due to the incorporation of multiple less relevant solutions, secondly, the feature map-based design methods ignore the characteristics of the feature vectors themselves, and thirdly, they lack a comprehensive consideration of the semantic relationships of the categories. In this paper, a brain-inspired modelling framework is proposed to model the neural circuits of hippocampal. The TSP module simulates situational memory, uncovering image properties, by interacting with pairs of sample features, while the MSP module simulates statistical learning, forming overlapping representations, by optimizing the spatial distribution of the samples' feature states. The work in this paper integrates the semantic information embedded in feature vectors to generate discriminative representations in terms of their universal nature. Experimental results show that the model in this paper is effective in public benchmarks and can be a complementary alternative to other state-of-the-art methods.