Low dimensional CNN (LDCNN)
This is the implementaton of the paper "Learning low dimensional convolutional neural networks for high-resolution remote sensing image retrieval ".
LDCNN is a CNN architecture consisting of common convolutional layers and the mlpconv layer used in Network in Network (NIN). The features extracted by the pre-trained CNNs like AlexNet, VGG, etc., from the fully-connected layers are usually 4096-D, while LDCNN can learn low dimensional features and generally have better performance than the pre-trained CNNs. The matlab implementation is available on GitHub!
We also provide some finetuned CNNs (CaffeRef, VGGM, VGGS) and the trained LDCNN to replicate the results in Table 10. The finetuned CNNs and LDCNN are trained on AID dataset and then evaluated on UCMD, RSD and RSSCN7 datasets.
If you use this software, please cite the following work:
Zhou, W., Newsam, S., Li, C., & Shao, Z. (2017). Learning low dimensional convolutional neural networks for high-resolution remote sensing image retrieval. Remote Sensing, 9(5), 489.