基于目标形状卷积神经网络在舰船分类中的应用

针对传统卷积神经网络采用通用卷积核提取目标特征造成更高的时间和空间开销的问题，提出一种适应目标几何形状的卷积核结构以替代通用卷积核，可使单个卷积核充分提取目标特征，简化目标提取过程，减少冗余计算。实验以网上收集的舰船可见光图像数据集为研究对象，实验结果表明:本方法在舰船目标识别任务中达到了99.7%的分类准确率，与目前通用的分类模型进行对比要高出约1%，训练速度是通用模型中收敛速度最快的模型的3倍。

Application of Convolution Neural Network Based on Target Shape in Ships and Warships Classification

The traditional convolution neural network uses general convolution kernel to extract target features， which results in significant time and space overhead. In this paper， a convolution kernel structure suitable for target geometry was proposed to replace the general convolution kernel， which can make a single convolution kernel extract more comprehensive target features. Additionally， it can simplify target extraction process and reduce redundant calculation. The experiment takes the visible image data set collected from the internet of ships and warships as the research object. The experimental results show that the method achieves a classification accuracy of 99.7% in the ships and warships target recognition task， which is about 1% higher than that of existing general classification models， while the training speed of our model is 3 times faster than that of state-of-the-art general models.