Object Recognition with Sequential Decision Reinforcement of Deep Learning

Abstract

The great success of deep learning methods for object detection rendered such methods the fundamental choice in related applications. Popular choices for multiple object detection in video sequences include convolutional neural networks, such as YOLO, MobileNet-SSD and Faster R-CNN, which typically split image frames to small rectangular regions and attempts to find bounding boxes of sought–after objects. Current research of such methods mostly focus on speeding–up the implementations or improving the network layers’ learning properties. As a new approach, this work appends a simple post processing stage at the end of such networks to reinforce decision robustness using a sequential decision process through sequential video frames. The sequential frames provide a better confidence on the existence of an object, when a probable object was also estimated in the previous frame. Once the confidence level overshoots a predetermined threshold, objects that are difficult to be detected in a single frame get accurately detected.