Object Recognition And Detection Based On Synthetic Images

Object recognition and detection occupy an important position in computer vision and have a wide range of applications in the military,security and civil fields.In recent years,with the development of deep learning technology,the recognition and detection performance has been improved significantly.However,the success of deep learning relies heavily on large-scale datasets,requiring a large number of images with annotations,which has caused difficulties in practical applications.In recent years,with the development of image technology,synthetic images are gradually applied as the training data in deep learning.The application of synthetic images in object recognition and detection is studied as follows:First,a large number of synthetic images are obtained by rendering 3D models,and are evaluated on typical convolutional networks.Two deep learning models are used to classify and recognize objects.Some factors set in the rendering process such as illumination,rendering accuracy and background are experimentally analyzed to determine the best setting of parameters in the process of image generation,so that images of good quality can be used as the dataset.Second,classification models are designed to improve the recognition performance on the real images.Though the classification model is trained by synthetic images,the ultimate goal is to recognize real objects which the traditional models are poor at.So classification models are designed for two situations to improve the results.Firstly,a new generative adversarial network model is proposed when only synthetic and few real labeled images are available.The combination of adversarial loss and classification loss makes the discriminator of the model fully learn the characteristics of the two types of images which increases the accuracy by10%.Secondly,based on transfer learning and Image Net dataset,the network features are transferred by fine-tuning,deep adaptation network and deep co-adaptation network which can bridge the gap between the synthetic and real image domains.Thus,the recognition accuracy of real images increases to over 98%.Last,synthetic images are used in object detection tasks.The performance of synthetic data is studied under the Faster R-CNN object detection framework based on pre-trained model in previous work.The experiment results show that few real images mixing with synthetic images can relieve the impact of lack of labeled data.In addition,the competition mechanism is introduced into the convolutional neural network to reduce the sensitivity of the model to the unimportant details in the synthetic images.The experiment results show that the proposed method can further improve the detection performance.