Rise of smartphones, cheaper cameras and improved image recognition thanks to deep learning based approaches opened a new era for image recognition. Companies in different sectors such as automotive, gaming and e-commerce are adopting this technology.
While selecting an image recognition solution, its accuracy is the most important factor, however continuous learning, speed and flexibility can also be important criteria depending on the application. Tech giants like Amazon and Google as well as startups like Clarifai are offering image recognition services.
What is Image Recognition?
A specific object or objects in a picture can be distinguished by using image recognition techniques.
Image recognition is set of algorithms and techniques to label and classify the elements inside an image. Image recognition mainly focuses on contents inside an image. Image recognition models are trained to take an input image and outputs previously classified labels that defines the image. Image recognition technology is an imitation of the techniques that animals detect and classify objects.
How is image recognition different than computer/machine vision?
Although image recognition and computer/machine vision may appear to be interconnected terms, image recognition is a subset of computer vision.
Image recognition is a technique for identifying the content of an image.
Computer vision involves obtaining, describing and producing results according to the field of application. Image recognition can be considered as a component of computer vision software. Computer vision has more capabilities like event detection, learning, image reconstruction and object tracking.
Machine vision is the vision system involving both hardware and computer vision software. Therefore, computer vision and image recognition can be considered as components of machine vision software
How does image recognition work?
Model training is necessary for an image recognition model to work. Deep learning methods are currently the best performing tools to train image recognition models.
In order for an image recognition model to work, first there must be a data set. Consider a newborn baby, in order for the baby to identify the objects around him, the objects must first be introduced by his parents. The process is similar for machines, there is a data set and using deep learning techniques, the model must be trained in order to perform.
An image is a bunch of pixels to a computer. In order to make a meaningful result from this data, it is necessary to extract certain features from the image. This process is called feature extraction. Feature extraction allows specific patterns to be represented by specific vectors. Deep learning methods are also used to determine the boundary range of these vectors. At this point, a data set is used to train the model, and in the end the model predicts certain objects and labels the new input image into a certain class.
Why is Image recognition software relevant now?
Image recognition allows meaningful data to be extracted from an image and therefore has numerous applications. However, image recognition accuracy only recently improved, making image recognition relevant in the past ~10 years. These factors contributed to increased usefulness of image recognition:
- Increased effectiveness of deep learning: We explained how deep learning became more powerful in the last ten years in detail.
- Reducing camera size and cost, coupled with increased smartphone penetration and image based social media: Thanks to the smartphone technology and photo-sharing social media platforms, images are proliferating. According to MarketsandMarkets “image recognition market is estimated to grow from USD 15.9 billion in 2016 to USD 38.9 billion by 2021, at the CAGR of 19.5% during the forecast period.”
What are image recognition use cases?
The technology behind the self driving cars are highly dependent on image recognition. Multiple video cameras and LIDAR create the images and image recognition software help computer to detect traffic lights, vehicles or other objects.
The ability to detect and identify faces is a useful option provided by image recognition technology. Home security systems are getting smarter and more powerful than they used to be.
Detecting tumors or brain strokes and helping visually impaired people are some of the use cases of image recognition in healthcare sector. A research shows that using image recognition, algorithm detects lung cancers with 97 percent accuracy.
Thanks to image recognition technology, Topshop and Timberland uses virtual mirror technology to help customers to see what the clothes look like without wearing them.
Visual search and e-commerce
An article at Forbes states that visual search market is expected to exceed $25 billion by 2021. The main reason is visual search is integrated with online shopping and customer habits are changing on this way.
Image data in social networks and other media can be analyzed to understand customer preferences. This data can be used for customized marketing for example. A Gartner survey suggests that image recognition technology can increase sales productivity by gathering information about customer and detecting trends in product placement.
What are the things to pay attention to while choosing Image recognition solutions?
The most important criteria for an image recognition solution is its accuracy, i.e. how well it identifies images. Criteria like speed and flexibility come later in most applications.
We explained in detail how companies should evaluate machine learning solutions. Once a company has labelled data to use as a test data set, they can compare different solutions as we explained. In most cases, solutions that are trained using companies own data are superior to off-the-shelf pre-trained solutions. However, if the required level of accuracy can be met with a pre-trained solutions, companies may choose not to bear the cost of having a custom model built.
Users should also not rush to make generalizations based on a single test. A vendor that performs well for face recognition may not be the appropriate vendor for a vehicle identification solution because the effectiveness of an image recognition solution depends on the specific application.
Other criteria include:
- Continuous learning: Every AI vendor boasts of continuous learning but few achieve it. The ideal solution should be learning from its incorrect predictions (inferences in machine learning jargon)
- Speed:The solution must be fast enough for the necessary application. While a customer facing solution may require a response within milliseconds, a solution for internal use can be OK to be produced within a few hours or even days.
- Adaptability for the future needs: It is important to foresee the constraints of the future and adaptability of the solution for the future needs is important.
- Simplicity of setup and integration: The solution should be easy to setup and use. Since most solutions will be API end points, they tend to be easy-to-setup.
What are image recognition software vendors?
Creating a data set and a neural network model and training it from scratch is not the most efficient way to take advantage of image recognition technology. Tech giants and some startup companies offer APIs that allow anyone to integrate their image recognition software. There are also open source APIs that can be used to build or improve your image recognition system. Feel free to browse our sortable list of leading image recognition providers.
There are also industry specific vendors. For example, Visenze provides solutions for visual search, product tagging and recommendation. You can learn more about Visenze by reading our interview here.
- Amazon Rekognition
- Google Cloud Vision API
- IBM Watson Visual Recognition
- Azure Face API
- Emotion API
- Computer Vision API
- Video API
If you have questions about how image recognition consultants can help your business, we can help:
How can we do better?
Your feedback is valuable. We will do our best to improve our work based on it.