针对海面背景舰船目标单一波段图像识别率低的问题,提出了一种基于卷积神经网络(CNN)的融合识别方法。该方法提取可见光、中波红外和长波红外3个波段舰船目标特征进行融合识别。模型主要分为3个步骤:通过设计的6层CNN,同时对三波段图像进行特征提取;利用基于互信息的特征选择方法对串联的三波段特征向量按照重要性进行排序,并按照图像清晰度评价指标选取固定长度的特征向量作为目标识别依据;通过额外的2个全连接层和输出层进行回归训练。采用自建的三波段舰船图像数据库进行模型的训练和测试,共包含6类目标,5000余张图像。实验结果表明,本文方法识别率达到84.5%,与单波段识别方法相比有明显提升。

In order to improve the recognition rate in single-band images of ship targets with complex background, we propose a new fusion recognition method based on convolutional neural networks (CNN). This method extracts the ship target features of images in three wave bands, which are visible light, medium-wave infrared and long-wave infrared images. The model is divided into three steps. Firstly, a 6-layer CNN model is designed to extract the image features of three bands simultaneously. Secondly, a feature selection method based on mutual information is used for sorting the concatenated features according to the importance, and then the feature vectors of fixed dimension can be chosen depending on the indicator of image clarity evaluation. The dimension-reduced feature vector is regarded as the basis of target recognition. Finally, a 2-layer fully connected networks and an output layer are designed for training and regression. We build a triple-band ship target dataset for our experimental verification, which contains 6 categories of targets and more than 5000 images. The experimental results show that the recognition rate of the proposed method can reach 84.5%, which is improved significantly compared to that of the single-band recognition method.