Research On Image Mosaic Based On Deep Learning And Reconfigurable Processor

In recent years,with the improvement of computer computing ability and the current demand for image quality,the human demand for wide-angle and high-quality images is improved.As a key technology of computer vision,image mosaic technology has been widely used in remote sensing photography,panoramic automobile and other fields.Deep learning has achieved good results in image classification,image recognition,target detection and other image fields.Based on the structure characteristics of deep learning network and traditional image mosaic methods,this thesis proposes an image mosaic method based on image semantic feature points extraction.The convolutional neural network model is used to quantify the semantic features of pixels in the image,and image mosaic is carried out according to the feature points obtained from the quantitative analysis.Compared with the shallow neural network,the semantic feature extraction method proposed in this thesis can reduce the dependence on the shallow features of the image.Based on the traditional image quality assessment method and deep learning method,this thesis proposes a new method of image quality assessment based on convolutional neural network.Before training and using the network,we first need to preprocess the image.By using image normalization and image segmentation,the image is segmented into two models.Then,when training,the pictures are matched with the matching of large and small pictures into a convolutional neural network for training.Finally,all the input images of each image segmentation are evaluated,and the quality of the original image is evaluated by using the quality mean of all the segmentation images as the way of the original image quality.The accuracy of the distortion type on the LIVE data set is 98%,and the Spearman order correlation coefficient and the Pearson linear correlation coefficient both reach 0.95.Based on the research of image mosaic and image quality assessment technology based on deep learning,this thesis maps and implements the deep learning basic operator by using reconfigurable processor.Reconfigurable processor is a processor consists of universal control processor and reconfigurable hardware.It combines the programmability of GPP with the high computing speed of ASIC,and is more suitable for processing large data than the traditional general-purpose processor.The processor can improve the operation efficiency of the operator greatly by configuring algorithms.Especially in the field of image processing and deep learning,reconfigurable processors can maximum their advantages because of their flexible configuration and good parallel computing ability.It can perform such algorithms well,which increases computing speed by 10 times compared to traditional ATOM processors.