One of the most representative problems in computer vision & machine learning that has attracted most of the attention from the research and engineering community is the accurate detection and segmentation of the objects in images. In current AI-based systems, the object detection problem is performed by providing the coordinates of the corresponding Region Of Interest (ROI) enclosing the detected object. On the other hand, a more accurate identification of the objects within an image, involves the classification of the objects at pixel level. This is what we call semantic segmentation, where the pixels of the image are assigned to different object classes.
The accurate detection and segmentation of the objects of interest within an image allows to build more intelligent and autonomous solutions for a wide range of applications in many different fields such as engineering, robotics, medicine, etc. The accurate detection & segmentation of the objects within an image can be of special interest to provide higher levels of knowledge to the system. For example, to infer objects dimensions, state, and some other features to predict future behavior of the objects. In some applications, such as infrastructures inspection, the AI-based system needs to determine the state of the detected object (i.e., whether the detected object is in good condition or presents some anomaly). In these applications, the IA-based models for object detection and segmentation are critical since posterior diagnosis models can be affected by the previous detection & segmentation result.
Current AI-based solutions for object detection & segmentation are usually based on Artificial Neural Networks (ANNs) which are trained on meaningful Datasets of images. Within the Supervised Learning paradigm, these datasets are usually labelled by a human operator, providing the source of ground truth from where the AI-based system is able to learn. The most common type of ANNs models used for object detection & segmentation are based on Convolutional Neural Networks (CNNs), which are capable of learning the appropriate convolution filters during the training process, to extract meaningful features from the images. These convolutional layers can be stacked and connected in many ways to build different neural network architectures. At Unusuals we develop cutting-edge AI-based systems having at the core ANNs and image processing algorithms with the purpose of automating the inspection and diagnosis process of different infrastructure assets.
