Research On Free-Hand Sketch Recognition Based On Convolution Neural Network

Free-hand sketches have been used as a common communication method since ancient times.In recent years,they have attracted widespread attention with the development of electronic devices and touch screens.They have also caused a series of studies on free-hand sketches,such as free-hand sketch recognition,free-hand sketch retrieval and 3D model retrieval based on free-hand sketches,etc.Free-hand sketch recognition is the focus of many scholars at present,and is also the basis of this field.Free-hand sketches is difficult,because free-hand sketches have many unique features compared to traditional natural images,which make them more challenging to identify than natural images.For example,free-hand sketches are drawn by non-professionals,different people have different painting characteristics and style,which makes the free-hand sketches highly abstract and exaggerated.In addition,the free-hand sketches are all generated by black lines,without corresponding color and texture information,which also makes the neural network can not learn the corresponding features when performing feature learning,which makes the image recognition more difficult and affects the recognition accuracy.Although there are many features of free-hand sketches that affect its recognition accuracy,it has its own advantages.For example,each free-hand sketch contains stroke order information,and the current dataset includes this information,which makes it easier for us to use this feature improves recognition accuracy.In the image recognition task of deep learning,constructing a network with high recognition rate requires a good network structure and uses a lot of data for network training.But many existing dataset is small,which can not meet the requirements of a large amount of data when training the network,and when the training data set is too small,the network is easy to produce over-fitting and affect the performance of the network.There are a lot of data enhancement techniques,such as basic flip,rotation,random cutting,and noise addition.These enhancement techniques can be used to increase the training data and improve the generalization ability of the training model.In addition,for special datasets and images,we can use special data enhancement techniques to increase the datasets.When using convolutional neural networks for image processing,the biggest problem is that the time cost of training a network is too high,training a complete network with better performance usually takes several days or even months,it is a big performance bottleneck fora realistic application.In our free-hand image recognition,it also takes a long time to train a network with good performance.With the emergence of a large number of high-speed computing hardware and the development of high-performance computing,we can try to apply parallel computing to CNN training based on CNN's natural parallelism.Common parallel technologies are OpenMP,MPI,CUDA,etc.They provide interfaces and platforms for parallel computing and are the most widely used parallel methods.In this paper,a network structure DCSN designed for free-hand sketch recognition according to the characteristics of free-hand sketches.DCSN network is improved based on AlexNet network,such as using a larger convolution kernel in the first layer to obtain the structural information of the sketch,using smaller steps in the first layer to retain the feature information,and increasing the network depth by increasing the number of network layers,etc.These improvements effectively improves the recognition accuracy of the free-hand sketch.In order to solve the over-fitting problem of the network and further improve the recognition accuracy,this paper proposes two data enhancement strategies,the small graphics reduction strategy and the tail removal strategy.In the training process of convolutional neural network,we found that the training speed of the network is slow,and it takes a long time to train a network.In order to reduce the training time of convolutional neural network,this paper uses a method that parallel training convolutional neural network to train the DCSN network,this method uses OpenMP and MPI hybrid programming model,making full use of the advantages of OpenMP intra-node shared memory and MPI can transfer data between nodes,complete two-level parallel between intra-node and nodes and nodes,effective reduce the time for convolutional neural network training.