《表1 REM网络结构：CCNet:面向多光谱图像的高速船只检测级联卷积神经网络（英文）》

  提示：宽带有限、当前游客访问压缩模式

本系列图表出处文件名：随高清版一同展现

《CCNet:面向多光谱图像的高速船只检测级联卷积神经网络（英文）》

1. 获取 高清版本忘记账户？点击这里登录

1. 下载图表忘记账户？点击这里登录

The detailed network configuration and receptive field（RF）of each layer in the REM are listed in Table1.Each convolution layer is followed by a ReLU active function.The feature map size of the output of conv4 is1/24 of the size of the original REM input image.And the receptive field of conv4 feature maps can reach 53×53.Each point of the conv4 represents a 53×53 image patch with stride 24.We use the central 32×32 pixels of the REM input image as the mapping of a point in conv4feature maps.Then the softmax is used to each point in conv4 feature map.The result of softmax represents the category of the corresponding 32×32 region in a layer of the image pyramid.The region of the original image which is represented by the 32×32 region in a layer of the image pyramid is captured as the ROI，such as a 32×32 region in the top of the 5-level image pyramid represents a 512×512 image patch of the original image.The ROIs generated in different layers of the image pyramid will be screened by using non-maximum suppression（NMS）method.If the adjacent image patches are classified as ships，they will be merged into an entire region as the final ROI.