Research on few-shot object detection algorithm

Inhalt:
Aiming at the situation that there are few labeled data samples in the neural network, an attention mechanism-based AM-SiameseRPN++ Few-shot Object Detection (FSOD) algorithm is proposed. The AM-SiameseRPN++ model extracts image features by fused residual network, and fuses different features of the multi-layer feature layers extracted by ResNet-50 through the feature fusion module. At the same time, an attention mechanism is introduced in front of the RPN network. The SiameseRPN++ and the AM-SiameseRPN++ algorithm proposed in this paper are used for feature extraction experiments on the Fashion-MNIST dataset. The experimental results show that AM-SiameseRPN++ has better feature expression and feature extraction capabilities than SiameseRPN++. The experimental results of 1-shot, 5-shot and 10-shot FSOD of different categories on the MS-COCO dataset show that AM-SiameseRPN++ model is more stable, has a certain inhibitory effect on the impact of object category growth, and has more advantageous FSOD performance. Therefore, AM-SiameseRPN++ is an effective FSOD algorithm.