基于卷积神经网络的云雾遮挡舰船目标识别

Obscured ship target recognition based on convolutional neural network

海上舰船目标识别对于海运交通、海上目标跟踪、军事侦察等都有着重要作用,然而海面气象复杂、光照不均、云雾遮挡等自然现象易导致遥感图像中舰船目标识别率低、鲁棒性差等问题.针对云雾遮挡问题,提出一种改进InceptionV3网络模型InceptionV3-FC的舰船目标识别算法.首先, InceptionV3-FC通过引入一层全连接层用来学习新的目标函数,用该目标函数对清晰样本和遮挡样本进行训练;其次,通过目标函数的约束项对清晰样本以及遮挡样本的特征进行约束,进而使得训练的遮挡样本和清晰样本的特征映射彼此接近,共享它们的特征,提高被遮挡舰船目标识别的鲁棒性.由光学遥感图像数据集的实验可知,相比于改进前的网络,将被云雾遮挡达到30%、50%、70%的舰船目标平均识别率分别提高3.23%、4.44%、15.67%.实验结果表明,该网络模型能有效改善舰船被云雾遮挡后特征丢失而导致的识别率低的问题.

The technology of target recognition plays an important role in maritime traffic, maritime target tracking and military reconnaissance, etc. The complex ocean ambient results in the information of ship targets is incomplete, so the cloud obscured ship target will have low recognition rate and poor robustness. Therefore, this paper proposes a obscured ship target recognition algorithm based on the convolutional neural network. First, the improved Inception V3 network is used to learn a new objective function, which is used to train the clear samples and occluded samples. Then in order to share the features of obscured samples and the clear samples, a constraint function is added to the loss function. Finally,experiments are performed on the optical remote sensing image dataset. The proposed method can improve the average recognition rates of the obscured ship target by 3.23 %, 4.44 %, and 15.67 % than the unimproved network which are obscured 30 %, 50 %, 70 % respectively. The experimental results show that the network model can effectively improve the low ship target recognition rate caused by cloud obscured.