联合正交特征融合与大核可分离注意力的道路分割算法

汤毅杰,钟铭恩,袁彬淦,范康,谭佳威,林志强

Road segmentation algorithm based on joint orthogonal feature fusion and large kernel separable attention

Yijie TANG,Mingen ZHONG,Bingan YUAN,Kang FAN,Jiawei TAN,Zhiqiang LIN

表 1 不同算法的性能对比结果

Tab.1 Performance comparison results of different algorithms

算法名称 Massachusetts数据集 DeepGlobe数据集 参数量/106 FLOPs/109 FPS/(帧∙s−1) P/% R/% IoU/% APLS/% P/% R/% IoU/% APLS/% U-Net[18] 77.29 72.13 59.46 66.84 73.20 70.10 59.06 65.37 26.4 223.9 158 SegNet[19] 72.79 77.41 60.11 66.97 79.84 75.92 63.79 65.69 29.5 170.5 133 Deeplabv3+[20] 75.47 77.97 62.25 67.76 78.20 76.24 62.33 65.86 54.7 83.2 93 D-LinkNet[21] 74.57 78.85 61.75 67.98 73.50 81.38 63.36 66.05 31.1 33.6 145 SDUNet 77.56 74.57 61.34 68.05 78.40 80.43 65.91 67.84 80.2 353.3 50 DSCNet 75.83 77.47 62.22 70.18 77.03 75.91 62.76 70.59 4.5 40.4 211 RoadExNet[22] 82.46 72.89 63.10 68.86 77.76 77.14 63.51 69.35 31.1 33.8 138 OARENet[23] 77.79 75.23 61.96 67.05 79.88 76.70 64.04 69.84 71.3 99.9 64 FRCFNet[24] 76.59 76.84 63.47 67.34 77.48 78.73 65.09 67.11 12.3 37.8 166 CFRNet[25] 77.63 77.06 63.77 72.29 78.93 78.89 66.28 71.51 36.6 35.7 169 UNetMamba[26] 77.07 78.34 63.89 73.15 78.43 79.38 66.22 72.28 14.8 25.1 171 LCMorph 76.93 78.51 63.55 — 78.51 81.02 66.30 — 71.9 294.6 — LightFormer[27] 77.52 78.65 64.21 73.91 78.86 81.33 66.79 74.42 13.7 23.9 172 ARSNet(ours) 78.12 79.63 64.85 74.95 79.78 81.63 67.88 75.53 22.5 14.8 178