Development is one of the oldest traits humanity possesses. It is encoded in our genome to always strive to improve, to grow, to make everything better. In 2020, during this era of technology evolution and gathering new ideas is the key to success. There is a single field that is showing great potential in radically changing the world and that is artificial intelligence.
Currently we are at the starting point of this journey, working with narrow AI technology, the one that is only responsible for a specific task and nothing more. To move forward the technology has to be altered to be as human as possible, before it reaches the stage of artificial superintelligence. Inspired by the human brain, artificial neural networks were created, marking the beginning of machine learning.
One of the methods of the machine learning family is deep learning, also known as deep structured learning. It is a technique developed to imitate human intelligence where machines can master things by experience and gain knowledge and skills without human intervention, learning from large amounts of data. The deep neural network analyzes data in a similar way a person would do it: the algorithm is given raw data and has to decide by itself which elements are relevant and worth proceeding with and which are better left out. Based on received feedback, it can learn the hierarchy of items automatically from data using a multiphase framework that represents a multi-level view ranging from pixels to high-level semantic features. Just as the human brain, these learning networks improve exponentially with the amount of data one trains them with, providing the user with better results every time.
The word ‘deep’ in the term indicates that the neural networks have many complex layers that can give the resources for learning. Similarly to the human brain, this machine learning branch is able to perform any task that requires ‘thought’ to solve a problem. The popularity of deep learning increased in 2006 following the surprising results regarding speech recognition. According to computer engineers, this method leads to significant advancements in areas such as object detection, face and speech recognition, drug discovery, image super-resolution reconstruction and medical imaging by allowing systems to learn complex yet subtle representations. However, it is worth mentioning that the results are only as good as the data the algorithm is trained on. If the data consists of unintentional bias or issues with integrity, machine learning can mislead the user.
Another state-of-the-art technology that is rightfully gaining a lot of importance these days is object detection. It is a combination of image classification and accurate object localisation based on deep learning that grants a proper and complete understanding of the image. The technology is related to computer vision, the science of understanding and manipulating digital images as if they were perceived by the human brain. Various application domains are being used for different tasks.
Salient object detection uses segmentation on pixel-level and local contrast improvements while generic object detection is using bounding box regression (BB) for detection, which is a popular technique to improve or predict location of the object by creating a rectangular box around it. Society has been benefiting from artificial intelligence while completing numerous daily tasks. Currently object recognition technology is widely used in surveillance, to control criminal behaviour, military and transportation, along with medical diagnosis and autonomous driving. Precise analysis of remote sensing images is very valuable for agricultural fields and defense industry. By feeding the algorithm with classified images one can train the program to accurately identify specific objects with a low error rate.
Needless to say, AI-based medical devices are being praised by their promising accuracy. Trained to match input images to the collected data library, this technology helps dealing with a massive amount of information, solving problems better and faster. It is successfully being used in detecting diabetic retinopathy with a precision of more than 87.4% as well as analysing X-rays and computed tomography scans. The current state of epidemics in the world only increases the need for this technology. Object tracking is another area of image processing to be achieved by unique identification and tracking the positions of identified objects over time. This field would prove itself useful in tracing the targeted moving pedestrians or vehicles, accident, criminal or security monitoring, although the technology is still in progress.
One of the tools produced merely for working with image recognition is SentiSight.ai. It is a web-based platform that is used for image labeling in addition to developing AI-based image recognition applications. The platform is developed by Neurotechnology, a company that values high-precision algorithms and software based on deep learning, and other AI-related technologies. Before the era of innovations, image commentary was carried out by particular team members responsible for this task. This method was costly, time and man-power consuming.
SentiSight.ai was created having only two objectives in mind: to make image labeling tasks more efficient and advantageous even when working on large scale projects and to provide users with a smooth and intuitive interface for training and using deep neural network models, saving companies precious time and money. The model is convenient for beginners as it is easy to use, as well as professionals who expect many training options, detailed statistics and beneficial results representation. It offers diverse performance tools, such as image similarity search, picture labeling, a smart labeling tool which considerably increases the speed of bitmap annotating, classification model training that helps to identify certain items in the picture, and object detection, predicting their exact location. There is a feature allowing numerous users to work on the same project simultaneously while being able to track time spent doing specific tasks. SentiSight.ai can be used either online, via the REST API server or web interface, or offline via downloaded local version for customers that want to run models on their own machines. The offline model can be used as a free trial for 30 days, however, after the time period ends it is vital to acquire a licence to continue to use it. The price depends on required speed for the server, ranging from 100 to 2000 EUR. More information about the licence and subscription pricing can be found here. Furthermore, SentiSight.ai provides several pre-trained models that can be used without any supplementary training for tasks such as goods classification and people counting.
As artificial intelligence technology continues on evolving, the main objective for us is to find innovative ways to use it for the benefit of society. At the moment there is a lot of room for improvement, although with enough research and trial and testing we could be entering the era of general AI soon.