为解决距离选通式激光成像中由于图像模糊而导致目标识别率偏低的难题,提出一种保留特征的卷积神经网络(KFCNN)模型,用于激光选通图像中的目标识别。与传统的卷积神经网络不同,KFCNN使用一个特征保留层来提高模糊目标的识别率,提高目标识别的稳健性。为实现特征保留,KFCNN通过增加特征保留约束项及正则化来优化特征保留目标函数并进行训练,通过减小特征保留目标函数值来保证训练样本在模糊之前和之后的特征映射相一致。实验结果表明,KFCNN改善了因模糊造成识别率降低的问题,进而提升了距离选通式激光成像中对指定目标的识别率。

In order to solve the difficult problem of low target recognition rate caused by image blur in the laser rang-gated imaging process, we propose the keep-feature convolutional neural network (KFCNN) model for the target recognition of laser rang-gated images. Different from the convolutional neural network (CNN), the KFCNN model is used to improve the recognition rate of blurred targets and the robustness of target recognition with a new keep-feature layer. To achieve keep-feature in the KFCNN model, we optimize the keep-feature objective functions and the training by imposing keep-feature constraints and regularization. In addition, the feature maps of training samples are kept consistent before and after image blur when the value of the keep-feature objective function is reduced. The experimental results show that KFCNN improves the problem of recognition rate reduction caused by image blur and further improves the recognition rate of specified targets in laser rang-gated imaging.