2326wx/sharp-in

Improve image quality

Sharp-In

Improve image quality utility

Description

The tool is designed to improve the images quality. At the moment, the tool accepts 512x512 JPG RGB images and returns images of the same size.

The before/after results look as follows:

See more examples.

Usage

1. Clone this project.
2. Download weights to the sharp-in project folder.
3. Create the MyImages folder at the project folder.
4. Put files to predict to the MyImages folder and then run predict.py.
5. The resulting images will appear shortly in the MyImages folder.

Your own weights training

The tool was trained to improve photos from drones. If you want to train it for another domain, create your own dataset with the prepare_dataset_for_superresolution.ipynb notebook.

General idea of the training

You can use any images for training unless they meet the following requirements:

• 512x512 size images as the target data.
• The same images of the reduced quality as the training data.

The dataset preparation notebook works both with annotated and non-annotated images.

Training dataset preparation

From each image in the dataset_paths folders it:

1. Takes crop of initial_crop_size.
2. Resizes the crop to crop_size - this will be y images.
3. Compresses it with the compress_ratios ratios and reshapes back to output_size - this will be X images.
4. The resulting X and y images are stored in the respective folders inside crops_folder.

You can use the notebook with annotated dataset with two types of crops:

• To get crops around target areas, run get_target_crops().
• To get crops_per_image random crops from each source image, run get_random_crops().

Other parameters:

• b_files_per_dataset: qty of files in the folder used for target crops.
• r_files_per_dataset: qty of files in the folder used for random crops.

Training

When the dataset is prepared, run train.py to start the training.

The recommended settings to start with: Adam optimizer with lr=1e-3 and Cosine LR scheduler. The batch size of 8 or 16 is acceptable.