Removing something from a photo doesn't have to be a hassle anymore, see how we use GAN to do it
When there is an unwanted object or passerby in the photo, the first thing we always feel is to remove it with retouching software. This idea is ingrained in you and me, showing how important it is for everyone to remove unnecessary clutter to present important information.
In the field of aerial survey, we often need to stitch the collected drone or satellite photos into digital orthophoto maps or 3D models for subsequent application and analysis. In some cases, the objects in the photos, such as vehicles, pedestrians, and animals, are not essential information. In addition to making the picture too cluttered, it is also difficult to show the real situation underneath, such as obscured road markings, building features, etc. However, we can't block the entire mission area to clear the road when we are on a flight mission to collect data, let alone stop birds from flying in. This is where object removal plays a very important role.
Bottlenecks in traditional practices
Object removal is the process of erasing an object from the target area and replacing it with a similar appearance as if the object never existed.
The common traditional method is to use retouching software to manually remove objects one photo at a time, such as Photoshop is a very well-known software. However, this may seem simple, but there are many inconveniences.
1. Labor costs
Retouching relies on experienced skills, not everyone can easily get started, you need to hire professional retouching staff, and it costs a few thousand dollars a month to hire each person.
2. Equipment costs
Retouching requires a computer with more advanced equipment, and parts for image processing are expensive, making a computer assembled cost close to 30,00 or more. Together with the purchase of professional retouching software, the cumulative cost is very significant.
With a limited budget and limited manpower, it will take a lot of time and effort to get a clean and beautiful result, and it will be more troublesome to make adjustments in the future. But if we want to improve efficiency, we need to hire more professional staff and buy more computer equipment, which in turn increases the cost significantly, but we do not always need so many resources, so the overall benefit is not worth it.
4. Not apply in some situations
In the field of aerial survey, for example, the raw data collected in each mission are hundreds of photos, which would be an extremely difficult task if processed manually.
Using deep learning practices
Deep learning techniques have given better solutions, and even the built-in object removal features of Samsung and Google phones were developed using deep learning techniques. As mentioned in the article "GAN - Automatic Coloring of Black and White Images", how GAN can be applied to automatically colorize black and white photos, GAN can also be applied to object removal by using artificial intelligence to automatically detect objects in a photo and excavate their areas, and then use GAN techniques to fill in the gaps in the photo.
For the object removal function of aerial survey images, we have developed a suitable GAN model and an automated data pipeline in cooperation with ZeroDimension Technology with the following features.
1. Automatic operation, low cost, high efficiency
Using the object removal tool we expect to develop in our online data transformation platform, you can simply upload the original file and wait for the result to come out. The process in between is automated, eliminating the need to spend resources on underutilized manpower and equipment, and making it more efficient.
2. Abundant items that can be removed
Our trained models have been able to recognize up to 80 objects, from vehicles, pedestrians, airplanes, and signs, to small glasses, and small animals such as cats, dogs, sparrows, etc., and remove them.
3. High-resolution processing
To allow objects to be erased more precisely and avoid removal to other areas, we enhance the object area's resolution to sharpen the object's edges.
4. Small object identification
Our model is able to identify and remove very small objects from photos, effectively avoiding the problem of aerial photos where the objects are too small to be identified.
Take a look at our example.
The application of deep learning has substantially increased convenience and moreover solved the cost-effectiveness problem of traditional manual methods. In the future, you only need to upload data and you can get the data you want automatically. Not only for aerial surveys but also for object removal from images, which can be handled by our GAN object removal tool.
For image processing and data transformation, DataXquad meets all your needs!
DataXquad is a Pay-Per-Use online image data transformation service platform that can meet all procedures without use costs. We simplify the most complicated part of image data transformation so that more image data can be utilized and more industry chains can derive value from it.