Research On The Recognition Of Small Targets Based On Super-Resolution

The recognition of small targets has very important application value in the fields of security,medical image analysis,and autonomous driving.Therefore,the precise recognition of small targets has important theoretical and practical significance for analyzing various image data.In recent years,the recognition algorithms based on deep learning have made a great breakthrough.However,the recognition of small targets is still a less research in the field of deep learning.At present,most recognition networks do not apply to the recognition of small targets.Due to its small size of small targets,the deep network could not be directly used for the recognition of small targets,and the shallow network loses its excellent performance.Therefore,in view of above problems,the research on super-resolution reconstruction and the recognition of small targets based on deep learning has been carried out in this paper.Aiming at the problem of accessing the datasets of small targets,the datasets used for the task of super-resolution reconstruction and the recognition of small targets are built through manual data annotation and data preprocessing.All datasets above are built based on the DOTA dataset and the actual collected UAV dataset,which can provide sufficient data for the research in this paper.Aiming at the problem that small targets have low resolution,less information and poor image quality,a new algorithm for super-resolution of images(VDSRGAN)is proposed based on an improved generative adversarial network in this paper.On the one hand,the idea of residual learning is introduced into VDSRGAN,which learns the residual between super-resolution images and low-resolution images instead of directly learning super-resolution images in SRGAN,and such an idea alleviates the problem that training SRGAN network is difficult.On the other hand,the additional feature layers are added to the new algorithm,which could help the generator restore clearer,more detailed super-resolution images.A large number of comparative experimental results indicate that the new algorithm could achieves better performance both on the objective evaluation index and the visual quality of images.Aiming at the problem that the recognizers perform poorly in the recognition of small targets,a multi-task deep network(M-SRRENet)is proposed on the basis of VDSRGAN,which integrates the super-resolution reconstruction and the recognition of small targets.In order to guide the training process of M-SRRENet better and improve the recognition performance of M-SRRENet when recognizing small targets,we further propose a new loss function(MTG-Loss).And the particle swarm optimization(PSO)algorithm is applied to optimize the weight coefficient,which could help obtain the optimal weight coefficient value in the solution space.Quantities of comparative experiments and ablation experiments indicate that the new algorithm could effectively improve the recognition performance of different types of small targets.In theory,the new algorithm could also be applied to the recognition of other types of small targets not included in this paper.