Small Sample Image Classification Based on Lightweight Inverted Residual Network Hybrid Prototypical Networks

Small sample image classification is an important issue in the field of computer vision, and it has a wide application prospect in resource-limited edge devices. Therefore, this paper proposes a hybrid model based on lightweight inverted residual network and ProtoNet to realize efficient small sample image classification. First, we design a lightweight inverted residual network as a feature extractor. By using optimization techniques such as point-by-point convolution and depth separable convolution, the complexity and computation of the model are greatly reduced, while maintaining good feature expression ability. Next, we fuse the lightweight inverted residual network with the Prototypical Networks (ProtoNet) classifier. ProtoNet is a small sample classification algorithm based on metric learning. It can use a small number of training samples to efficiently learn category prototype vectors and perform well in small sample scenarios. By combining a lightweight feature extractor with a ProtoNet classifier, our hybrid model can significantly reduce model complexity and inference time while ensuring classification performance, making it ideal for deployment on resource-constrained edge devices. We have conducted extensive experimental evaluation of the model on the public small sample image classification dataset, and the results show that the hybrid model is superior to other mainstream small sample classification methods in classification accuracy, reasoning speed, parameter quantity and other indicators. In general, the hybrid model based on lightweight inverted residual network and ProtoNet proposed in this paper provides an efficient and feasible solution to the small sample image classification problem, which is of great significance to edge devices and practical applications to reduce computational complexity.