Deep Learning Based Detection And Recognition Of Pokers

Target detection and recognition is the most important step in computer vision tasks.The detection and identification of pokers are based on pokers,classification and position the pokers in an image.It has important application value in the process of realizing intelligent Doudizhu.The pokers in the Doudizhu scene will have a large number of categories and mutual occlusion of the targets.These affect the detection and recognition effects.In view of the characteristics of poker image data and the needs of practical applications,this paper takes deep learning,neural network and target detection as theoretical and technical guidance to study the detection and recognition of poker images.The main research contents are as follows:1.An image data set for detection and recognition has been established.In view of the fact that there is no data about pokers in the current target detection data set,a solution for making a special image data set for pokers is proposed.First of all,collect the original images in the game from the actual network fighting platform,which can verify the feasibility of applying the algorithm in actual scenarios;secondly,filter and sort the collected images,such as removing incomplete images,uniform size,etc.,A total number 4000 poker images were obtained,with a total of 79040 playing card targets.Then the target image processing is carried out on the image data of the pokers;finally,the labeled image data is divided into training,verification,testing and other parts according to a certain ratio,and then the image data of each part and the labeling data are packaged into a TFRecord format data file To get the training set,verification set,and test set.2.Design and implement the playing card recognition algorithm based on the deep learning framework.The deep separable convolution is used to modify the image feature extraction part of the SSD algorithm,which reduces the number of parameters of the traditional convolutional neural network and implements the algorithm N-SSD for identifying playing card targets.And use the aforementioned playing card image data set to train and verify the N-SSD network model.The experimental results show that the average detection accuracy of the N-SSD algorithm on the verification data set reaches 90%,and it has a relatively high detection accuracy.3.Based on the N-SSD playing card recognition algorithm and a B / S architecture,a prototype playing card recognition system is designed and implemented.First,export the trained poker model,freeze the parameters,and deploy the service;then use the client to remotely call the model on the server for target detection,and display the detection results on the client.At the same time,the API interface for playing card detection has been opened.Algorithm source code and experiment: https://github.com/yesyihua/N-SSD?Flask.