Since 2010, the field of computer vision and computational image analysis has been revolutionized by the advent of effective Deep Learning techniques.
These techniques are based on the use of neural networks which are substantially deeper than the existing few years ago. Depth of the network allows develop richer models, which can automatically learn, in an unsupervised way, the most relevant visual characteristics of the analyzed images. These characteristics depend, in general, on the type of image we are working on, as well as the specific task we are using (e.g. an image classification, features searching and object classification within an image, detecting visual similarity between two images, etc.).
In what Computer Vision research concerns, this step ahead has meant that open challenges can be solved, and, for the first time in history, computational algorithms can understand images with greater subtlety and accuracy than trained human beings, even in highly specialized fields such as health.
The Computer Vision team in TECNALIA, with a large experience in industrial applications, has extended the use of these techniques to final industrial applications, getting much more accurate and better results. One example is the surface quality inspection of bars with defects as cracks, roll marks on its surface which have to be detected. TECNALIA has introduced Deep Learning techniques for the identification of these defects whose detection is critical for the quality improvement of the final product in steel industry.
The techniques in the analysis process lets:
- Increasing the rate of detected defects
- Reducing the rate of false defect detection
Therefore, it is possible to generate highly robust and accurate solutions for industrial companies were we have already increased drastically the accuracy of the detection systems up to 99%. Deep Learning can also be implemented in other sectors and for different parts and assemblies.