Journal of Communication and Computer vol:10 issue:12 pages:1484-1496
Day-to-day industrial computer vision applications focusing on object detection have the need of robust, fast and accurate object detection techniques. However, current state-of-the-art object categorization techniques only reach about 85% detection rate when performing in the wild detections, which try to cope with as much scene and object variation as possible. However, several industrial applications show many known characteristics like constant lighting, known camera position, constant background ..., giving lead to several constraints on the actual algorithms. With a complete new universal object categorization framework, we want to prove the detection rate of these object categorization algorithms by exploiting the application specific knowledge which can help to reach a robust detector with detection rates of 99.9% or higher. We will use the same constraints to effectively reduce the number of false positive detections. Furthermore, we will introduce an innovative active learning system based on this application specific knowledge that will drastically reduce the amount of positive and negative training samples, leading to a shorter and more effective annotation and training phase.