基于深度卷积神经网络的红外场景理解算法

Infrared Scene Understanding Algorithm Based on Deep Convolutional Neural Network

采用深度学习的方法实现红外图像场景语义理解。首先,建立含有4类别前景目标和1个类别背景的用于语义分割研究的红外图像数据集。其次,以深度卷积神经网络为基础,结合条件随机场后处理优化模型,搭建端到端的红外语义分割算法框架并进行训练。最后,在可见光和红外测试集上对算法框架的输出结果进行评估分析。实验结果表明,采用深度学习的方法对红外图像进行语义分割能实现图像的像素级分类,并获得较高的预测精度。从而可以获得红外图像中景物的形状、种类、位置分布等信息,实现红外场景的语义理解。

We adopt a deep learning method to implement a semantic infrared image scene understanding. First, we build an infrared image dataset for the semantic segmentation research, consisting of four foreground object classes and one background class. Second, we build an end-to-end infrared semantic segmentation framework based on a deep convolutional neural network connected to a conditional random field refined model. Then, we train the model. Finally, we evaluate and analyze the outputs of the algorithm framework from both the visible and infrared datasets. Qualitatively, it is feasible to adopt a deep learning method to classify infrared images on a pixel level, and the predicted accuracy is satisfactory. We can obtain the features, classes, and positions of the objects in an infrared image to understand the infrared scene semantically.