Efficient Scene Text Detection in Images with Network Pruning and Knowledge Distillation

Abstract

Modern deep learning models draw a progressive graph of success in many different areas. However, enrichment of these models and increase in their success rates depend on the systems which need high computing resources. In addition, the need for more computational power limits the real life usage of these models trained with deep learning. Scene Text Detection is one of the most popular topics, which must work under these limitations. To this end, network pruning and knowledge distillation, which are popular methods for network compression, were implemented with a text detection system, whereby faster, lighter and more accurate models are obtained. Moreover, training strategies and their details of these models are shared in this paper. The results of comprehensive experiments and model outputs demonstrate the success of the methods.