基于词向量一致性融合的遥感场景零样本分类方法 下载: 958次
Zero-Shot Classification Method for Remote-Sensing Scenes Based on Word Vector Consistent Fusion
1 海军航空大学, 山东 烟台 264001
2 空军航空大学, 吉林 长春 130022
3 91977部队, 北京 102200
图 & 表
图 1. 本文方法的整体框架图
Fig. 1. Whole framework of proposed method
下载图片 查看原文
图 2. 本文方法运算流程图
Fig. 2. Operational flow chart of proposed method
下载图片 查看原文
图 3. UCM数据集若干类的样本。(a)农田;(b)飞机;(c)棒球场;(d)密集住宅;(e)高速公路;(f)海港;(g)储罐;(h)网球场;(i)立交桥;(j)高尔夫球场
Fig. 3. Images of several classes from UCM dataset. (a) Agricultural; (b) airplane; (c) baseball diamond; (d) dense residential; (e) freeway; (f) harbor; (g) storage tanks; (h) tennis court; (i) overpass; (j) golf course
下载图片 查看原文
图 4. AID数据集若干类的样本。(a)机场;(b)贫瘠地;(c)海滩;(d)桥梁; (e)商业区;(f)运动场;(g)池塘;(h)火车站;(i)体育场;(j)立交桥
Fig. 4. Images of several classes from AID dataset. (a) Airport; (b) bare land; (c) beach; (d) bridge; (e) commercial; (f) playground; (g) pond; (h) railway station; (i) stadium; (j) viaduct
下载图片 查看原文
图 5. RSSCN7数据集类的样本。(a)草地;(b)河湖;(c)工厂;(d)场地;(e)森林;(f)居民区;(g)停车场
Fig. 5. Images of several classes from RSSCN7 dataset. (a) Grass; (b) river laker; (c) industrial; (d) field; (e) forest; (f) residential; (g) parking
下载图片 查看原文
图 6. 不同模型词向量融合的结构对齐效果
Fig. 6. Structure alignment performance of word vector fusion with different models
下载图片 查看原文
图 7. 不同语料词向量融合的结构对齐效果图
Fig. 7. Structure alignment performance of word vector fusion with different corpora
下载图片 查看原文
图 8. UCM和AID数据集上本文方法在不同α值上的OA值。(a) UCM上不同训练模型词向量融合;(b) UCM上不同训练语料词向量融合;(c) AID上不同训练模型词向量融合;(d) AID上不同训练语料词向量融合
Fig. 8. OA values of proposed method for different α on UCM and AID datasets. (a) Fusion of word vectors from different training models on UCM dataset; (b) fusion of word vectors from different training corpora on UCM dataset; (c) fusion of word vectors from different training models on AID dataset; (d) fusion of word vectors from different training corpora on AID dataset
下载图片 查看原文
图 9. 不同训练模型词向量和不同训练语料词向量的各unseen类融合效果。(a) UCM上不同训练模型词向量融合;(b) UCM上不同训练语料词向量融合;(c) AID上不同训练模型词向量融合;(d) AID上不同训练语料词向量融合
Fig. 9. Fusion performance of different training models and different training corpora on unseen classes. (a) Fusion of word vectors from different training models on UCM dataset; (b) fusion of word vectors from different training corpora on UCM dataset; (c) fusion of word vectors from different training models on AID dataset; (d) fusion of word vectors from different training corpora on AID dataset
下载图片 查看原文
图 10. 本文方法及对比方法的各unseen类S1+S2+S3词向量融合效果。 (a) UCM数据集; (b) AID数据集
Fig. 10. Fusion performance of S1+S2+S3 word vectors on unseen classes by proposed method and relative methods. (a) UCM dataset; (b) AID dataset
下载图片 查看原文
图 11. 测试遥感图像I的场景ZSC效果图
Fig. 11. Scene ZSC results of test remote-sensing image I
下载图片 查看原文
图 12. 测试遥感图像II的场景ZSC效果图
Fig. 12. Scene ZSC results of test remote-sensing image II
下载图片 查看原文
表 1不同训练模型词向量和不同训练语料词向量融合前后的OA
Table1. OA values of different training models and different training corpora before and after fusion of word vectors%
Dataset | Fusion of word vectors from different models | Fusion of word vectors from different corpora |
---|
gl | wv | glwv | | Wiki | Crawl | WikiCrawl |
---|
UCM | 51.39 | 47.65 | 61.23 | 51.39 | 42.86/44.61 | 59.77 | AID | 57.92 | 61.53 | 69.47 | 57.92 | 56.16/58.29 | 68.49 |
|
查看原文
表 2本文方法及对比方法OA值
Table2. OA values of proposed method and relative methods%
Method | UCM | AID |
---|
S1 | S2 | S3 | S1+S2 | S2+S3 | S1+S2+S3 | | S1 | S2 | S3 | S1+S2 | S2+S3 | S1+S2+S3 |
---|
LatEm[4] | 18.80 | 20.40 | 19.80 | 33.00 | 23.00 | 20.80 | 15.90 | 22.65 | 23.81 | 18.71 | 28.17 | 21.62 | RKT[5] | 40.00 | 39.80 | 44.60 | 40.20 | 43.60 | 43.60 | 48.92 | 48.03 | 48.15 | 48.92 | 50.13 | 53.25 | DMaP[9] | 38.20 | 39.60 | 41.60 | 40.80 | 42.00 | 40.20 | 39.24 | 43.44 | 38.54 | 46.67 | 45.22 | 44.97 | BiDiLEL[8] | 28.51 | 33.48 | 39.20 | 40.40 | 40.00 | 41.00 | 32.91 | 42.55 | 32.40 | 47.85 | 50.44 | 49.63 | JLSE[6] | 37.25 | 34.21 | 45.68 | 37.66 | 34.88 | 38.03 | 36.11 | 34.97 | 42.30 | 35.99 | 43.50 | 45.54 | SSE[7] | 38.36 | 39.48 | 37.91 | 38.72 | 39.19 | 38.23 | 38.24 | 37.16 | 39.56 | 34.53 | 40.92 | 43.56 | Proposed | 44.61 | 47.65 | 51.39 | 61.16 | 61.23 | 68.56 | 58.29 | 61.53 | 57.92 | 70.44 | 69.47 | 76.85 |
|
查看原文
表 3各ZSC算法在AID数据集上对S1词向量上的运算耗时
Table3. Computing time of different ZSC algorithms on AID dataset with S1 word vector
Method | Time/s |
---|
LatEm[4] | 21.66 | RKT[5] | 24.24 | DMaP[9] | 409.26 | BiDiLEL[8] | 28.81 | JLSE[6] | 70.20 | SSE[7] | 19.74 | Proposed | 17.90 |
|
查看原文
吴晨, 于光, 张凤晶, 刘宇, 袁昱纬, 全吉成. 基于词向量一致性融合的遥感场景零样本分类方法[J]. 光学学报, 2019, 39(8): 0828002. Chen Wu, Guang Yu, Fengjing Zhang, Yu Liu, Yuwei Yuan, Jicheng Quan. Zero-Shot Classification Method for Remote-Sensing Scenes Based on Word Vector Consistent Fusion[J]. Acta Optica Sinica, 2019, 39(8): 0828002.