Feasibility Study On Application Of Simulated Images In Deep Learning

At present,compared with ordinary machine learning algorithm,deep learning algorithm has higher research value,but it also faces unavoidable problems in its application: deep learning relies on the support of large-scale data,and the performance of the model trained by a small number of samples is often low.In some cases,it is difficult to obtain large-scale samples.For example,the image data in infrared scene,in fact,the cost of image acquisition is high,and it is difficult to obtain data under various conditions.Therefore,for image processing in this field,the application of deep learning algorithm seems to be inadequate.The large-scale image data can be acquired more quickly by image simulation method,and various scene factors can be added according to the need.Therefore,the use of simulation image as the data set of deep learning algorithm needs further study.Because there are some differences between the simulated image and the real image,the feasibility of using the simulated image as the data set of the depth learning algorithm is discussed in this paper.Starting with the task of target recognition and detection,this paper chooses the target field which is scarce and difficult to obtain by real-time photography,namely,the ship target in infrared scene,and carries out image simulation,and uses a large number of samples obtained from simulation to train the convolution neural network model algorithm.In terms of algorithm,the existing mature deep learning target recognition and detection algorithm,SSD algorithm,is selected for testing.In this paper,the small amount of real-time infrared warship data is obtained as a small number of samples through Internet resources,which verifies the performance of the small number of sample training methods under the deep learning algorithm.The results are compared by simulating a large number of samples,and the feasibility of the simulation image as a deep learning data set is illustrated.Finally,although there are some problems in using a small number of samples to directly train the convolutional neural network model,the small number of sample data as real data has greater value.Therefore,this paper improves the process of training by using simulation images only.Based on the idea of Tr-Ada Boost algorithm,the weight of the small number of real samples and the large number of simulation samples in the training process is allocated.Finally,it is found that the improved training model has higher prediction performance,and the reasonable method of using simulation image as deep learning data set is pointed out.