A NOVEL APPROACH FOR OBJECT IDENTIFICATION WITH DEEP REINFORCEMENT MACHINE LEARNING

This paper presents an active model with convolutional neural net concept for object identification which uses deep reinforcement learning by applying action learned by agent finds position of object and with the help of Softmax classification fully connected layer defines the class of object. The agent decide with the policy how to place a bounding box over object using eleven action and different transformation, Comparing to existing technique our method uses less computation and also finds objects in minimum steps. For getting object features we create shallower F-RCNN with three convolutions and two fully connected layer which acts as pretrained model it provide the object features and with combination of this information RL will find the object location and place bounding box over object. Softmax Classifier used to define class information. Implementation of proposed model have done with anaconda 3 (64bit) with tensor flow which shows the better results comparing to other technique in literature.