摘要
卷积神经网络在图像识别方面具有独特的优越性,但在实际场景中,其识别结果会受到图像背景的干扰。船舶识别中图像的背景因素极其复杂,因此,结合显著性检测算法,从图像中分离出待识别的船舶,通过卷积神经网络进行船舶识别。鉴于显著性检测很难在复杂图像中完整的分离背景和前景,提出两种改进卷积神经网络的方法,即"中心-扩散池化"卷积神经网络和"前景-扩散池化"卷积神经网络。实验结果表明,改进的卷积神经网络表现出更稳定的表征能力和更好的泛化能力,结合显著性检测算法改进的卷积神经网络在船舶识别中取得了很好的成效。
Convolution neural network has unique superiority in image identification,but in the actual scene,the recognition result is disturbed by image background.The background factor of the image in ship identification is extremely complex.Therefore,the saliency region detection algorithm was combined to separate the ship to be identified from the image,and the ship identification was carried on through the convolution neural network.In view of the fact that the saliency region detection is difficult to separate the background and the prospect in the complex images,two methods of improving the convolution neural network were proposed,namely center-diffusion pooling convolution neural network and foreground-diffusion pooling convolution neural network.Experimental results show that the improved convolution neural network exhibits more stable characterization ability and better generalization ability.Combining saliency region detection algorithm,the improved convolution neural network method achieves good results in ship identification.
作者
杨亚东
王晓峰
潘静静
YANG Ya-dong;WANG Xiao-feng;PAN Jing-jing(College of Information Engineering,Shanghai Maritime University,Shanghai 201306,China;College of Transportation and Civil Engineering,Fujian Agriculture and Forestry University,Fuzhou 350002,China)
出处
《计算机工程与设计》
北大核心
2018年第10期3228-3233,共6页
Computer Engineering and Design
基金
国家自然科学基金项目(31170952)
国家海洋局基金项目(201305026)
上海海事大学研究生创新基金项目(2017ycx083)
关键词
显著性检测
卷积神经网络
中心-扩散池化
前景-扩散池化
船舶识别
saliency region detection
convolution neural network
center-diffusion pooling
foreground-diffusion pooling
ship identification
