Object Recognition And Detection Based On Light Field Image Super-resolution

With the large-scale popularization of imaging equipment,imaging activities can be carried out everywhere in our daily life.The massive dynamic and static image information collected by the imaging system provides abundant information resources for relevant researches in many fields,such as public security,production and life,military,and other fields.However,it is difficult for eye recognition methods to search the interest information quickly and accurately because of an enormous amount of image information,so image recognition technology has been developed and applied.However,the obtained images have poor quality and fuzzy information due to the limitation of the imaging device and the uncertainty and contingency factors of imaging scenes without human intervention,thus the target recognition is affected to some extent.Therefore,it is of great significance to carry out the object recognition method based on image super-resolution reconstruction research from the application value of images.The essential point of object recognition and detection method based on image super-resolution reconstruction is image super-resolution reconstruction technology,so the paper studies the process and premise of image super-resolution reconstruction.Given the complexity of traditional multi-image super-resolution reconstruction and the advantages of light field imaging,the light field image super-resolution reconstruction is adopted in our paper.In the research work,we use the light field camera to collect datasets and process these datasets to obtain the available light field multi-view image and then use the proposed method of object recognition and detection based on light field image super-resolution reconstruction to complete the research of the input image.The performance of the object recognition is improved based on enhancing the input image information.This paper mainly completes the following work:Firstly,a pedestrian recognition method based on the residual network light field image super-resolution reconstruction is proposed.According to the characteristics of multi-view images of the light field,there are sub-pixel shifts in different directions among views.Namely,fuzzy information in one view may be clearer in adjacent views.Therefore,the adjacent views of the reconstructed view are stacked,and then the superresolution reconstruction of the target view is achieved when the local and global feature extraction between views is completed.Finally,the pedestrian recognition research of the reconstructed image is completed by Retina Net.The existence of residual blocks in both light field image super-resolution reconstruction based on the residual network and the Retina Net network makes the enhancement and recognition of the input image more compatible.Secondly,inspired by the above research work,we further proposed object detection based on the light field image super-resolution reconstruction method.Since detection requires higher image information contrast recognition,so the superresolution reconstruction based on the residual network for object detection is not applicable(the inadaptability is mainly caused by ringing artifacts in super-resolution image reconstructed images).In response to the above problem,a geometry-consistent light field super-resolution via graph-based regularization for object detection method is proposed.According to the research on the structure of the light field in geometryconsistent light field super-resolution via graph-based regularization,not only can the intensity of light field be extended in the whole field,but also the geometric consistency of the structure of the light field is maintained,and the reconstruction of image superresolution is transformed into a global optimization problem,which avoids the traditional sub-pixel displacement calculation problem.Then,the method of object detection is researched by using the reconstructed image.Experimental results show that our method is superior to others,especially the robustness of the method.This method is not only suitable for object recognition and detection,but also provides a new idea for challenging tasks in the future.