The human brain can clearly and easily recognize objects and boundaries in images. This skill enables us to look at any picture and easily determine whether the picture contains a car for example. We can not only determine whether a car is in the picture, but if it is, we can point to the exact location of the car in the picture. Image classification and object detection are prevalent in a wide variety of industries and applications. The primary task of airport security agents is to match faces in real life with ID cards and to detect unwanted objects in baggage. Inspectors on manufacturing lines are checking whether the image they see on the assembly line is a match for the design image. Automatic image detection and classification attempts to mimic the image-processing functionality of the human brain to develop efficient computer vision systems. The end goal is to enable computers to gain a high-level understanding from processing digital images. Image classification is the process of choosing a label to assign to new images. These labels are determined from the training data, so any image that we want to classify must have some example of the label we want in the training data. Most image classification models are restricted to subject-specific fields, so only labels relevant to that field are included in the training data. A tool to label and classify images of wildlife tracks will need training data that includes a variety of paw prints but wont ever label an image as an airplane or a car. General image classification models exist, but performance at specific tasks is improved with curated training data. Object detection is like image classification, but rather than assigning a label to the whole image, a box is drawn somewhere in the image and a label is assigned to that box. Because we need to draw these bounding boxes in addition to assigning labels, object detection is a harder task than image classification. Just like with image classification, training data must include labeled images, which in this case are images with bounding boxes drawn on each labeled object. Object detection is important for applications such as self-driving cars, where the car must detect the road boundaries, other cars, and obstructions in the road. In this lesson, we discuss how to use convolutional neural networks to create image classification models. Object detection algorithms have been implemented in SAS Viya but are beyond the scope of this course.
