融合双通道特征学习机制的图像铅垂方向识别

Image vertical direction recognition based on double channel feature learning mechanism

为准确识别图像的铅垂方向(即图像中物理环境的铅垂方向,简称为IVD),提出一种融合双通道特征学习机制的网络模型。该模型由三个模块组成,分别是数据预处理模块、特征学习模块以及IVD识别模块。首先,为全面地学习图像的旋转不变特征,在数据预处理模块中通过随机旋转组合的方式增广数据集。然后,在特征学习模块中构建包含特征复用通道和特征生成通道的“双通道”,在抑制特征退化的同时学习新特征,并利用通道注意力模块突出重要特征。最后,在IVD识别模块中构建全新的旋转损失函数,用以同时识别物体类别与判定IVD的偏离角度。此外,根据ILSVRC2012创建了一个可用于模型训练和测试的数据集。基于该数据集的实验结果表明,所提方法的分类精度优于主流神经网络方法,达到97.68%,且旋转损失函数能使模型的角度均方误差至少降低8.05%,证明了对图像内容的正确识别可以帮助模型更有效地学习图像的旋转不变特征。

In order to efficiently identify the vertical direction of an image(that is,the vertical direction of the physical environment in the image,referred to as IVD),a double channel feature learning model is proposed.The model consists of three modules:data preprocessing module,feature learning module and IVD recognition module.Firstly,the data set is augmented by random rotation combination process in the data preprocessing module to learn rotation‑invariant features more comprehensively.Then,a“double channel”process containing feature reuse channel and feature creation channel is constructed in the feature learning module to learn new features while suppressing feature degradation,and the channel attention module is used to highlight important features.Finally,a new rotation loss function is constructed in the IVD recognition module to identify the object category and determine the offset angle of IVD simultaneously.In addition,a data set is created for training and testing according to ILSVRC2012.Experimental results based on the dataset show that the recognition accuracy of the method is better than that of the popular convolutional neural networks and achieves 97.68%.The rotation loss function can reduce models’angle mean square error by at least 8.05%,which proves the correct recognition of image content can help the model learn images’rotation‑invariant features more effectively.