结构主导目标检测中的纹理随机化

已有基于卷积神经网络的目标检测算法倾向于提取目标纹理特征,而非结构特征;因此,已有方法不能实现变纹理目标的可靠检测。针对此问题,提出基于纹理随机化的结构主导目标检测方法,采用仿真纹理随机化方法减弱网络模型对纹理特征的拟合,实现基于结构特征的变纹理目标可靠检测。利用目标的三维模型,借助Blender渲染引擎,完成纹理随机化仿真训练数据集的生成。仿真及真实图像实验测试结果表明:该方法能够实现基于目标结构特征的变纹理目标可靠检测。

Texture randomization in structure-dominant object detection

Region-based detectors with convolutional neural networks tend to learn textural rather structural feature and thus face substantial difficulties in detecting objects with various textures. To tackle this problem, the texture randomization to augment the synthetic training image dataset was employed and a novel method for structure-aware object detection was proposed. The texture-randomized simulation data were generated by rendering 3D model with varied textures using Blender. Experiments on synthetic and real images indicate that the proposed method is capable of robustly detecting texture-varied objects based on structural information.