Optimized Reward Function Based Deep Reinforcement Learning Approach for Object Detection Applications

Reinforcement learning is considered a powerful artificial intelligence method that can be used to teach machines through interaction with the environment and learning from their mistakes. More and more applications are coming to the fore where Reinforcement learning has been newly and successfully implemented. It is frequently used especially in the game industry and robotics. In this article, a deep reinforcement learning approach, which uses our own developed neural network, is presented for object detection on the PASCAL Voc2012 dataset. Our approach is by moving a bounding box step-by-step towards the goal in order to fully frame the object in the picture. The created neural network consists of a 5-layer structure. In addition, it is aimed to maximize the mAP value by optimizing the reward function. The right choice in the reward policy will certainly affect the outcome and will play an important role in the training of the agent. Thanks to the optimized reward function, ground truth and the bounding box intersect at the highest rate, contributing positively to the result. As a result of the training that lasted for approximately 36 hours, the test results of 6 randomly selected classes were compared with the results of previous similar studies. Within the scope of this article, some artificial neural networks and basic studies in the literature using the Reinforcement learning approach for object detection are examined.