基于YOLOv5的伪装数据增广方法研究

Research on camouflaged data augmentation method based on YOLOv5

数据作为深度学习的驱动力,对于模型的训练至关重要。为解决自然伪装数据集样本数量不足的问题,提出一种基于纹理置乱的数据增广方法,可有效扩充数据集样本数量。首先,根据目标像素多少动态选择置乱尺度,然后,将目标区域按置乱尺度切分为若干纹理块,最后将这些纹理块的排列顺序进行置乱。实验证明,使用基于纹理空间置乱的数据增广方法生成的数据保留了模型学习伪装目标的语义特征,同时增加了样本的多样性。在YOLOv5s模型上使用扩充的数据集进行训练,模型的检测性能提升了2.4个百分点,与传统数据增广方法相比,所提方法取得了更好的效果。

As the driving force behind deep learning,data is crucial for model training.In order to address the issue of insufficient samples in natural camouflage datasets,a data augmentation method based on texture perturbation is proposed to effectively expand the sample size.Firstly,the perturbation scale is dynamically selected based on the number of target pixels.Then,the target region is divided into several texture blocks according to the perturbation scale,and the arrangement order of these texture blocks is perturbed.Experimental results demonstrate that the data generated using the texture‑based perturbation data augmentation method preserves the semantic features of the model learning camouflage targets,while increasing sample diversity.Training on the augmented dataset on YOLOv5s improves the detection performance of the model by 2.4 percentage.Compared with traditional data augmentation methods,the proposed approach achieves better results.