The Aerial Image Recognition Challenge revolves around sorting different landscape images ranging from beaches, lakes, forests, meadows and more.
The data comes from parts of the data set (NWPU-RESISC45) originally used in the paper Remote Sensing Image Scene Classification. This data set contains 45 categories while we only kept 13 out of them as a first preprocessing.
)
This solution to the Aerial Image Recognition Challenge was developed using Jupyter over the course of 2 months. The group members were KHATER Yara, POBELLE Marion and SERRE Gaëtan.
We first used an instance of the REsNet50 model that is already trained on the ImageNet dataset. All pre-trained models expect input images normalized in the same way, i.e. mini-batches of 3-channel RGB images of shape (3 x H x W), where H and W are expected to be at least 224. The images have to be loaded into a range of [0, 1] (in order to have small weights for small data entries) and then normalized using mean = [0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225] (both mean and standard deviation with pixels). So we made our images look the same way as the images ResNet50 was pre-trained on. We obtained stable results for our loss and accuracy.
)
As the pretrained ResNet50 model was performing well, we decided to recreate its architecture but without any pretrained weights. Once again we obtained great results on loss and accuracy.
)