A method based on Graph Attention NN for image embedding and classification
In this step the high resolution image shoud be cropped to small 224 * 224 image. so each image is converted to a batch of small images.
In this step all of the small images will pass through a backbone network such as ResNet50 which is used in this work to obtaion feature vectors for each small image.
By obtaining the features from next part now we should convert these small images to a graph. I did that by addng edge between those images which have a cosine similarity less than 0.2. After this part the each high resolution image is converted to a graph.
By converting each image to a graph now we can perform graph operation on the image. In this work the goal is to find an embedding for each image so I used the lables for the classification task and perfomed graph classification algorithm to obtain the final embedding.
The dataset is from FloodNet challenge that can be accessed from Here