Research On Image And Video Recognition Based On Single Node Photonic Reservoir Computing

Traditional computers have become very mature after decades of development and evolution.They have shown excellent performance in many fields.So it is considered to be the expansion of the human brain.However traditional computers are far from being as well as the human brain in many problems.Artificial neural networks(ANN)process information in a way that mimics the human brain.It is expected to break through the difficulties encountered by traditional computers.It shows great potential in many fields.ANN mainly include feedforward neural networks and recurrent neural networks.Because recurrent neural networks have memory characteristics.Therefore,recurrent neural networks are good at executing time-related tasks.However,recurrent neural networks have not been widely used due to the complex network structure,long training period,and slow convergence speed.Reservoir computing(RC)is one paradigm of recurrent neural network,which is an improvement on the traditional recurrent neural network.Only the output connection weights require to be trained,whereas the input connection weights and the internal connection weights are fixed randomly.The single-node RC further simplifies the structure because of containing only one physical node which is of great significance to the integration of ANN.Due to the concept of high-dimensional non-linear mapping is used.A simple linear regression algorithm can be employed to train the RC,which overcomes the difficult problem of recurrent neural network training.Based on the single-node RC paradigm,we use a semiconductor laser with self-delayed optical feedback as a single physical node to build a single-node photon RC(P-RC).The white chaos collected in the experiment after differential processing is used as the mask signal.We train output connection weights based on the ridge regression algorithm.We introduce the single-node P-RC to the handwritten digit recognition.The original image is firstly transferred into associated feature descriptors using the histogram of gradient(HOG)technique.We employ the winner-take-all decision strategy to determine the output results.After optimizing main hyperparameters of the P-RC,the simulation results show that the recognition accuracy can reach 99.3%.Based on the single-node P-RC built,we performed video recognition of human behavior.The original video is firstly transferred into associated feature descriptors using the HOG technique by each frame.In the output layer,the result is determined as the behavior class of the majority frames recognized.Finally,the correct rate of human behavior video recognition based on the single-node P-RC reached 98% after optimizing system hyperparameters.The excellent performance of single-node P-RC in image and video recognition for computationally intensive tasks is verified,which is of great significance to the research of ANN.