Squeeze-and-Residual-Excitation Aggregated Attention Network For Improved Super Resolution in Remote Sensing Imagery
Squeeze-and-Residual-Excitation Aggregated Attention Network improves super-resolution (SR) on remote-sensing imagery compared to other state-of-the-art attention-based SR models.
The model is built in PyTorch 1.8.1.
Super-resolution (SR) provides an excellent approach of improving applications related to remote-sensing imagery. The tasks like object detection, classification and segmentation can greatly benefit from high-performing SR models. Substantial research is being carried out in the field of SR for both natural and remote-sensing imagery where deep convolutional neural networks (CNN) have achieved tremendous improvements. Many deep CNNs employ the attention mechanism in their structure and one such mechanism is the Squeeze-and-Excitation (SE) channel attention. While SE block has shown performance gains in many models, there is still room for improvement in its structure, therefore, in this paper, we propose the Squeeze-and-Residual-Excitation (SRE) attention block. SRE improves upon the SE block by employing residual mechanism within its structure to provide performance gain in SR. Based on our SRE attention mechanism, we propose an enhanced SR framework that outperforms other attention-based deep SR models for two levels of resolution enhancement: 4x- and 8x-upsampling on two diverse remote sensing imagery datasets: Satellite Imagery Multi-Vehicles Dataset (SIMD) consisting of 5000 high-resolution remote sensing images, and Cars-Overhead-With-Context (COWC). Moreover, by using YoloV5 object detection model, we perform numerous experiments to validate the effectiveness of our proposed SR model for the task of object detection on SIMD.
The setup can be done through one of the following methods:
- A new environment can be created using the 'requirements.yml' file by executing the following command:
conda env create -f requirements.yml
- Or it can also be created through the given "conda_environment.txt" file.
conda create --name sre-aan --file conda_environment.txt
Use the created environment for testing and performing experiments.
Place images that you want to upsample in the 'demo/low_res_images' folder
CD to 'src' and run one of the following script
The upsampled images will be found in the 'demo/results/results-Demo' folder.
#for 4x upsampling
python main.py --template 4X_SRE_AAN --pre_train ../trained_models/sre_han_x4.pt --n_GPUs=2 --data_test Demo --dir_demo ../demo/low_res_images --test_only --save ../demo/results --save_results
#for 8x upsampling
python main.py --template 8X_SRE_AAN --pre_train ../trained_models/sre_han_x8.pt --n_GPUs=2 --data_test Demo --dir_demo ../demo/low_res_images --test_only --save ../demo/results --save_results
The super-resolution performance over the Satellite Imagery Multi-Vehicles Dataset of our model and three previous state-of-the-art models are shown in this section. Our model outperforms the other models for both 4x and 8x upsampling.
We use YoloV5-Medium mopel presented by Glenn Jocher to perform two types of object detection experiments.
YoloV5-Medium is trained and evaluated on the upsampled versions of the dataset obtained through using the super-resolution models. The results are as follows:
YoloV5-Medium is trained over the ground truth dataset and then evaluated on the upsampled versions of the test set obtained through using the SR models.
The visual results for the performance of the object detection model that has been trained on the ground truth dataset and then evaluated on the images upsampled through bicubic method and SRE-AAN model are shown here.
Ground truth trained YoloV5 Object Detection Model is applied over both of the above images and we can see that the detection model is unable to detection any of the vehicles in the bicubic upsampled image. On the other hand, Yolov5 detects many of the objects in the image upsampled through our SRE-AAN model.
Similarly, in the bicubic upsampled image the detection model detects only 1 object, whereas in the SRE-AAN upsampled image all of the objects have been detected except one.