Classify Images Without Labels

June 20, 2020, 4:26 p.m. By: Harshita Kaur

Image Classification without labelling

Image classification is a process that involves categorizing and labeling groups of pixels or vectors within an image based on some specific rules. Image classification is used in areas such as medicine, education, and security.

The two-step approach by decoupling the feature learning and clustering followed by the researchers of KU Leuven/ESAT-PSI and ETH Zurich/CVL, TRACE is different from the traditional one.

Their method of image classification involves assignment of semantic labels from a predefined set of classes to an image. The fundamental problem that gains attention in image classification is the determining of the category or class of an image. This problem becomes more complicated with the increase in the number of categories as several objects with different classes present in the image can belong to several categories considering the interest in semantic class hierarchy as well.

The images are labeled using c image annotation that have all the features and performs functions to annotate the images for different types of machine learning training, some of the image labeling techniques include semantic segmentation, bounding box, polygon annotation, landmarking annotation, cuboid annotation, polyline annotation.

In image annotation for machine learning, the perspective to easily recognize machines without any error is kept intact by labeling or annotating the image so that it becomes easy to recognize the objects and predict the result correctly. For instance: Self-Driving cars can detect pedestrians or any other moving object and drive in the right direction. This approach was applied to 1000 classes on ImageNet and the results were really good and encouraging.