Infrared Image Object Detection Based On GAN Data Augmentation

Object detection is one of the most important and challenging branches in the field of computer vision.It is widely used in many fields of image processing,including image retrieval,security supervision,automatic driving and industrial defect detection.With the rapid development of deep learning network for detection task,the performance of target detector has been greatly improved.The excellent performance of object detection model based on deep learning is inseparable from a large number of high-quality training datasets.However,in the field of infrared image,due to the limitation of infrared imaging equipment and the increase of labeling cost,there are few labeled datasets,so the application of convolutional neural network in thermal infrared image target detection is limited.Most of the existing datasets are focused on visible images,while thermal infrared images are helpful for target detection even in dark environment.In order to solve this problem,we use image to image style transfer models,which convert the available labeled visible images into infrared images.At the same time,we compare the traditional data enhancement model with Gan network data enhancement model,and use the contrast experiment to prove that the false infrared image datasets generated by Gan network has a positive effect in data enhancement.The average accuracy of the original datasets only using the traditional data enhancement method is 79.18%.The average accuracy of the original datasets can be improved to 82.24% by adding the generated image data.At present,in some application fields where computer vision algorithm has been put into practice,it has made a great breakthrough to solve the problem of insufficient precision on small datasets by using the excellent pretraining model in the industry as the initial weight.Therefore,the excellent pretraining model has become a breakthrough to solve practical problems.To solve this problem,we propose to use transfer learning to improve the accuracy of infrared pedestrian detection.We pretrain a convolutional neural network model on a large dataset(including 2.3 million images of 710 categories),and then finetune the feature extractor to complete the infrared pedestrian detection task on the basis of the convolutional neural network model parameters.Finally,the average detection accuracy of Image Net training model alone is 82.24%.The average detection accuracy of COCO training model alone is 82.51%.Using our pretraining model,the average accuracy is improved to 83.56%.We believe that this transfer learning method can be extended to other small dataset training applications,and make other breakthroughs.