An Effective Deep-Learning Training Method using Data Augmentation

Abstract

Deep learning is changing the research paradigm, showing dramatic performance improvements in many areas of computer vision. Methods/Statistical analysis: It should be < 70 words. Since Lecun’s Lenet was released, Deep Learning has achieved significant performance improvements in object recognition and classification. However, it takes a huge data and takes a long time in order to learn, making it difficult to apply to real industrial environments. This method requires many manpower, high know-how and a lot of development time.Therefore, we propose an effective deep-training training and performance enhancement method using data augmentation. After changing the original image to a YUV color space favorable to computer vision, the image is created by raising or lowering the luminance value in units of 5.Using the proposed data augmentation method can save time and cost.Findings: In order to achieve satisfactory performance by applying deep learning to a real industrial environment, we must useour own method of producing a huge amount of data. In addition, the method of producing a direct dataset requires collecting a large amount of image data setsfor a specific object and sorting the data with high quality.In this paper, we propose efficient learning method of SSD (Single Shot MultiBox Detector)deepening learning image object recognition model based on MobileNetwhich is widely used in a mobile environment and embedded environment and data augmentation method to improve recognition performance (mAP).Improvements/Applications: In SSDbased on MobileNet, the saturation of loss is faster than that of the original data set alone, and the mAP is improved by 0.7.