Deep reinforcement learning

The combination and cooperation between reinforcement learning and neural networks has been studied in this work. The advent of deep learning has made the application of reinforcement learning possible, especially in deep reinforcement learning (DRL). We utilize a deep deterministic policy gradient (DDPG) algorithm, which consists of an actor-critic method, where the actor is responsible for selecting proper policy and the critic is the reinforcement learning algorithm to learn an optimal policy. actor uses the policy gradient method, takes agent states as input variables, then outputs the compensate policy. The critic takes the compensate policy and reward as input variables, estimates the uncertainties, and outputs the gradient direction of the expected return, then updates the policy parameters in the actor.

 

 

Learning based image segmentation

To balance accuracy and real-time performance of image segmentation techniques, we propose a novel architecture, MFFM-Net for flood segmentation. The following figure shows an architecture of the proposed method, composed of MobileNet, MFF Block, and FPN. MobileNet has great real-time performance in feature extraction, but accuracy can hardly satisfy the requirement. To improve the real-time performance and accuracy of the network, we construct the MFF block. This block can help merge the context information of features better.