Object Detection Using Policy-Based Reinforcement Learning

By constructing an object detection model using a deep neural network structure, it was possible to obtain faster and more accurate results than traditional models. Afterwards, through the method of applying the transformer structure by replacing the existing convolutional layer, the model structure, which was previously carried out in two stages, was simplified, and it became possible to detect the size of the object more freely. However, there are difficulties in generating new data due to the complex configuration of data used for training. As a result, a lot of resources are consumed in data generation, and it is difficult to train in response to a new environment immediately. This study presents a method for learning using simpler data. The simplified data consisted of the input image and the number of objects to be detected. To train using the data, a reinforcement learning method was applied to evaluate the output of the detection model and create and train a reward based on this.