Modification of single-purpose CNN for creating multi-purpose CNN

Abstract

Modern traditional convolutional neural networks (CNN) are built to solve one specific task, which can be: detecting and localizing objects in an image, restoring a human skeletal model, classifying images, etc. However, existing business problems can be a combination of the above subtasks. The simplest way to solve this problem is to use several ready-made CNN models or to train such models, one for each subtask. But this approach leads to a strong increase in the required computing power, which can be economically costly. The approach presented in the article allows reducing computational costs by adding new blocks to the existing CNN and creating a multi-purpose model.