多尺度轻量化CNN在SAR图像地物分类中的应用

孙盛, 蒙芝敏, 胡忠文, 余旭

Application of multi-scale and lightweight CNN in SAR image-based surface feature classification

SUN Sheng, MENG Zhimin, HU Zhongwen, YU Xu

表1 ENet和ENet-CSPP的结构对比

Tab.1 Comparison of ENet and ENet-CSPP structures

ENet ENet-CSPP 层名 采样类型 输出大小/像素 层名 采样类型 输出大小/像素 initial 16×256×256 bottleneck1.0 downsampling 64×128×128 bottleneck1.0 downsampling 64×256×256 4bottlencek1.x 64×128×128 4bottlencek1.x 64×256×256 bottleneck2.0 downsampling 128×64×64 bottleneck2.0 downsampling 128×128×128 8bottleneck2.x 128×64×64 8bottleneck2.x 128×128×128 repeat section 2,without bottleneck2.0 CSPP N2×128×128 bottleneck4.0 upsampling 64×128×128 upsample1.0 upsampling N2×256×256 2bottleneck4.x 64×128×128 concat (N1+N2)×256×256 bottleneck5.0 upsampling 16×256×256 lastconv C×256×256 2bottleneck5.x 16×256×256 upsample2.0 upsampling C×512×512 fullconv C×512×512