Yolov3-Pruning(transfer): real-time object detection algorithm based on transfer learning

Abstract In recent years, object detection algorithms have achieved great success in the field of machine vision. To pursue accuracy model, scale network is constantly increasing, which leads to continuous increase computational cost and a large requirement for memory. The larger allows their execution take longer time, facing balance between speed execution. Therefore, developed algorithm not suitable real-time applications. improve performance small targets, we propose new method, based on transfer learning. Based baseline Yolov3 pruning done reduce then migration learning used ensure model. method using achieves good inference more conducive processing images. Through evaluation dataset voc2007 + 2012, experimental results show that parameters Yolov3-Pruning(transfer): model are reduced by 3X compared with improved, realizes processing, improves accuracy.