A Meta Learning Few Shot Object Detection Algorithm Based on PRN-FPN
Lianyuan Jiang, Jinlong Chen · 2024
The goal of Few Shot Object Detection (FSOD) task is to effectively train and deploy object detection models even in small datasets, in order to achieve high-precision object detection results. Due to the few shot size of the dataset, the model may have difficulty learning sufficiently generalized feature representations, which can lead to overfitting or underfitting issues. Specifically, the model performs well during training, but shows a significant poor performance on test. That's because model overfits the noise and specific samples of the data in the training set, resulting in the inability to generalize to new data. Secondly, in the case of few shots, the target category scale in the dataset is single, making it difficult for the model to distinguish the same target at different scales, which affects detection accuracy. In order to effectively perform FSOD tasks, this paper proposes a meta learning FSOD algorithm based on PRN-FPN. Firstly, a data augmentation module is proposed to address the issue of data scarcity in few shot situations. The module includes three conversion strategies: color, color, and histogram equalization. Combining these three strategies can effectively increase the diversity of scarce data. In order to improve the algorithm's recognition and detection of targets of the same category but different scales, a Predictor head Removing Network with Feature Pyramid Network (PRN FPN) module is proposed. The module adds CARAFE Feature Pyramid Network on the basis of Predictor head Removal Network (RPN), and has the ability to detect few shot objects at multiple scales by fusing features at different depths. Finally, the algorithm was trained and tested on the publicly available dataset Pascal VOC, and the experimental results showed its effectiveness.