Open image repositories for training AI models

Perhaps one of the most everyday uses of artificial intelligence that we can experience in our day-to-day lives is through interaction with artificial vision and object identification systems. From unlocking our smartphone to searching for images on the Internet. All these functionalities are possible thanks to artificial intelligence models in the field of image detection and classification. In this post we compile some of the most important open image repositories, thanks to which we have been able to train current image recognition models.

Introduction

Let's go back for a moment to late 2017, early 2018. The possibility of unlocking our smartphones with some kind of fingerprint reader has become widespread. With more or less success, most manufacturers had managed to include the biometric reader in their terminals. The unlocking time, the ease of use and the extra security provided were exceptional compared to the classic systems of passwords, patterns, etc. As has been the case since 2008, the undisputed leader in digital innovation in mobile terminals - Apple - revolutionised the market once again by incorporating an innovative unlocking system in the iPhone X using an image of our face. The so-called FaceID system scans our face to unlock the terminal in tenths of a second without having to use our hands. The probability of identity theft with this system was 1 in 1,000,000; 20 times more secure than its predecessor TouchID

Let this little story about an everyday functionality be used to introduce an important topic in the field of artificial intelligence, and in particular in the field of computer image processing: AI model training image repositories. We have talked a lot in this space about this field of artificial intelligence. A few months after the launch of FaceID, we published a post on AI, in which we mentioned near-human-level image classification as one of the most important achievements of AI in recent years. This would not be possible without the availability of open banks of annotated images^[1]  to train image recognition and classification models. In this post we list some of the most important (freely available) image repositories for model training.

Of course, recognising the number plate of a vehicle at the entrance to a car park is not the same as identifying a lung disease in an X-ray image. The banks of annotated images are as varied as the potential AI applications they enable.

Probably the 2 best known image repositories are MNIST and ImageNET.

• MNIST, is a set of 70,000 black and white images of handwritten numbers normalised in size, ready to train number recognition algorithms. Professor LeCun's original paper is from 1998.

• ImageNET is a huge database of concepts (words or sets of words). Each concept with its own meaning is called a synset. Each synset is represented by hundreds or thousands of images. ImageNET's own website cites the project as an indispensable tool for the recent advancement of Deep Learning and computer vision.

The project has been instrumental in advancing computer vision and deep learning research. The data is available for free to researchers for non-commercial use

The most widely used subset of ImageNet is ImageNet Large Scale Visual Recognition Challenge ILSVRC, an image classification and localisation dataset. This image subset was used from 2010 to 2017 for the worldwide object detection and image classification competitions. This dataset covers 1000 object classes and contains more than one million training images, 50,000 validation images and 100,000 test images. This subset is available in Kaggle.

In addition to these two classic repositories that are already part of the history of image processing by artificial intelligence, we have some more current and varied thematic repositories. Here are some examples:

• The very annoying CAPTCHAs and reCAPTCHAs that we find on a multitude of websites to verify that we are human trying to access are a good example of artificial intelligence applied to the field of security. Of course, CAPTCHAs also need their own repository to check how effective they are in preventing unwanted access. We recommend reading this interesting article about the history of these web browsing companions.

• As we have seen several times in the past, one of the most promising applications of AI in the field of imaging is to assist physicians in diagnosing diseases from a medical imaging test (X-ray, CT scan, etc.). To make this a reality, there is no shortage of efforts to collect, annotate and make available to the research community repositories of quality, anonymised medical images to train models for detecting objects, shapes and patterns that may reveal a possible disease. Breast cancer are 30% of all cancers in women worldwide. Hence the importance of having image banks that facilitate the training of specific models.

• The diagnosis of blood-based diseases often involves the identification and characterisation of patient blood samples. Automated methods (using medical imaging) to detect and classify blood cell subtypes have important medical applications.

• Three years ago, Covid19 burst into our lives, turning developed societies upside down with this global pandemic with terrible consequences in terms of human and economic loss. The entire scientific community threw itself into finding a solution in record time to tackle the consequences of the new coronavirus. Many efforts were made to improve the diagnosis of the disease. Some techniques relied on AI-assisted image analysis. At the same time, health authorities incorporated a new element in our daily routine - face masks. Even today, in some situations the mask is still mandatory, and during these 3 years we have had to monitor its proper use in almost all kinds of places. So much so that in recent months there has been a proliferation of specific image banks to train AI and computer vision models to detect the use of masks autonomously.

• For more information on open repositories related to health and wellbeing, we leave you with this post we published a few months ago.

In addition to these curious examples cited in this post, we encourage you to explore Kaggle's section of datasets that include images as data. You only have 10,000 sets to browse through ;)

Content prepared by Alejandro Alija, expert in Digital Transformation and Innovation.

The contents and views expressed in this publication are the sole responsibility of the author.