Robustness adversarial training to improve small object recognition
鲁棒性对抗性训练对小物体识别的提高

摘要

提出了一种基于对抗性训练的小物体识别方法，用于提高模型在复杂环境下的鲁棒性。该方法使用YOLOv3作为基础框架，冻结前几层并只更新后续层。通过真实图像和生成的复杂图像进行训练，使模型适应两种数据。采用细致的Adam优化器、较小的学习率和批大小进行训练。实验结果显示，该方法在小物体数据集上的mAP高于YOLOv3，并在复杂测试集上具有高精度，表明对抗训练确实增强了模型的鲁棒性。然而，该方法的速度下降到YOLOv3的0.7倍，因为对抗图像较复杂，需要更长的前向传播时间。总之，对抗性训练可以显著提高小物体识别模型的鲁棒性，但也会带来速度下降和数据集依赖性增加的问题。需要进一步改进模型和训练策略，以在保持鲁棒性的同时尽量减少速度和数据集影响。综上所述，该研究提出了一种基于YOLOv3和对抗性训练的小物体识别方法，可以显著提高模型在复杂环境下的鲁棒性，但还需要进一步改进和优化。

Abstract

A small object recognition method based on adversarial training is proposed to improve the robustness of the model in complex environments. The method uses YOLOv3 as the base framework, freezing the first few layers and updating only the subsequent layers. The model is trained through real images and generated complex images to adapt to both types of data. Training with a detailed Adam optimizer, small learning rate and batch size. Experimental results show that the proposed method has a higher mAP on small object datasets than YOLOv3, and has high accuracy on complex test sets, indicating that adversarial training does enhance the robustness of the model. However, the speed of this method drops to 0.7 times that of YOLOv3 because the counter image is more complex and requires a longer forward propagation time. In conclusion, adversarial training can significantly improve the robustness of small object recognition models, but it also brings the problem of reduced speed and increased dataset dependency. Further improvements to the model and training strategies are needed to minimize speed and data set impact while maintaining robustness. In summary, this study proposes a small object recognition method based on YOLOv3 and adversarial training, which can significantly improve the robustness of the model in complex environments, but it needs further improvement and optimization.