Recent Trends in Small Object Detection

Recently, we have been conducting research on small object detection in high-resolution imagery.

Although we specialize in high-resolution imagery, detecting an object that is clearly visible when zoomed in and detecting an object that remains small even after zooming (meaning that is shows poor resolution)  are totally DIFFERENT problems. For this second case, we have to rely on what is called Small Object Detection, and here we present some recent trends.

Recent attempts have been made to identify small objects by using super resolution techniques. Super-resolution imaging (SR) consists of a class of techniques that enhance (or increase) the resolution of an imaging system.

These techniques are not necessary when you have a powerful camera and storage system that can achieve very high resolutions. However, when you need to find distant (or small) objects in photos that are limited in terms of resolution given the hardware that is available and image capture conditions (probably when using a low-cost camera), then you can consider applying these methods.

Certainly, some of our users in Deep Block are experiencing this problem! ☺

Now, I'll explain this approach in slightly more detail, but since our mission is to enable people who are not engineers to use deep learning, we will focus on this at a conceptual level. This discussion is largely based on Small-Object Detection in Remote Sensing Images with End-to-End Edge-Enhanced GAN and Object Detector Network by Rabbi, et al. where you can read the technical details in depth.  

Let’s continue,  my friends.

It's actually a quite simple concept: a small object is detected in two steps.

1. Put the image in a super-resolution model to improve the resolution. This is also referred to as upscaling.

2. Then put the upscaled image into a traditional object detection model (like FRCNN) to detect the small objects.

Here is a very simple diagram of the steps:

Recent trend of small object detection-2-2

 

In this way, the overall process really is quite simple. The underlying implementation for each of the steps is quite difficult and takes a significant amount of engineering and computing time to train the models, which could be painful. Then, aside from the image processing, there can be many technical details to provide real-time serving, model optimization, additional data preprocessing capabilities, etc. 

But don't worry! This is where Deep Block and our team comes in. We take care of all the technical details under the hood, and you'd be able to focus on your work at a much more simple conceptual level.

If you, or someone you know, needs to get intelligent results with small objects from your image data, please contact us any time at https://www.deepblock.net/contact.

We are also building the Deep Block AI Suite for self-service image analytics, which you can try out today.

Please visit https://www.deepblock.net/