Research And Application Of Image Restoration Based On Attention

In the current big data era,the camera-based device allows people to capture a large number of digital image data.However,during the acquisition process,due to the influence of bad weather,light conditions and the hardware error of the imaging device,the quality of acquired image has different degrees of degradation,such as imaging content blur,low resolution,low brightness and contrast,etc.affecting needs of visual sensory,and even affecting the effects of actual visual tasks.Therefore,image restoration technology has important research and practical significance.In recent years,with the significant advancement of computer image processing chip technology,the deep neural network has been applied to various fields of image processing and has achieved excellent results,while the attention mechanism is a strong innovative part in the deep neural network.By analyzing the demand for practical applications and the problem of existing algorithms,in-depth exploration,analysis and research on the deep neural network image restoration algorithm based on attention mechanisms,three algorithms are proposed and their engineering applications are operated.The main research results of this paper are:(1)An image recovery algorithm based on physical model constraints and attention mechanisms is proposed.Most existing image restoration algorithms are based on the constructed image degraded physical model,and the key variables in the physical model are estimated with the prior or learning-based approach to achieve restoration.However,the variable range in the physical model is usually limited,and it is difficult to accurately estimate,while multi-step restoration process will result in an error accumulation.The method of directly learning the features of the original image through the end-to-end neural network is too dependent on the dataset,and it is difficult to adapt to the complex real scenes after being isolated from the physical model.For the above problems,the algorithm explores a new way of using image degradation physical models.First,the physical model is added as an additional constraint condition to the generating adversarial network,then the relatively clear image restored from the generating network is input into the physical model to generate a secondary degraded image,and discriminate it with the original degraded image.The model does not need to estimate the variables in the physical model,effectively reducing the error accumulation,and is not isolated from the physical model,making it more adapted to complex scene data.In addition,it is designed with the generating network using multi-level grid structure and channel attention mechanism,and the focus on the region that has strong relationship with the place in severe degradation is increased through self-attention module,better utilizing global information generation results.The experiments on the synthetic image dataset and the natural image dataset prove that this method is better than most mainstream algorithms,and the ablation experiments also prove the necessity and effectiveness of the core structure of the method.(2)A multi-stage image restoration algorithm based on non-local attention network is proposed.Multi-stage neural network has better effects than a single-stage neural network.However,existing multi-stage image restoration algorithms are ether difficult to keep spatial image details,or it is difficult to output semantic information.At the same time,existing methods lack the identification of information in the feature map,and few methods that take the information in the feature map into consideration is not ideal.In response to the above problem,the algorithm performs image restoration issues in three stages.The first two stages use the encoder-decoder structure to learn multi-level context semantics information,and in the highest stage,a non-local attention module is used to extract the features of the image with original resolution for keeping fine spatial details.In order to allow the network to take the association between the pixel blocks in the feature map into account,the combination of the non-local attention module and the deep network enables the network to capture the global information of the feature map,and by calculating the self-similarity,the feature map can be adaptively recalibrated.At the same time,a supervised attention module is inserted between every two stages to achieve progressive study.Under the guidance of the tag of true value,the module utilizes the output of the former stage to calculate the attention map,and then the attention map will be in turn act to adjust and improve the feature map of the former stage,passing to the next stage.Through the mechanism of feature fusion across stages,the network is enabled to transmit multi-level context semantic feature from the former stage to the latter stage.The experimental results show that both the subjective and objective effect of the algorithm has a certain degree of improvement than the current algorithm,especially obvious for the restoration effects of the image with rich texture.(3)The engineering application of the image restoration algorithm on the object detection task.In order to verify the role of the image restoration on the high-lever visual task,and compare the effectiveness of the image restoration algorithm proposed in this paper,this paper considers dangerous goods detection in the elevator as the basic application scenarios,on the basis of the two-stage hybrid concatenated object detection algorithm design,implementation of an engineering application of image restoration on the object detection task is achieved.Due to the influence of light,noise,distortion,etc.in elevators,the resolution of data is decreased,and identifying dangerous goods becomes more difficult.Therefore,it is necessary to use the image restoration algorithm as the preprocess of the object detection,and the identification accuracy of the object detection algorithm is improved by lifting the data quality.Aiming at the problem of diverse resolutions of images collected in community elevator,the image restoration algorithm across multi resolutions is designed.Hence,the dangerous goods detection framework for low-quality image in elevator car is constructed by experimentally verifying the effectiveness of the image recovery algorithm on the self-built dataset for dangerous goods detection in elevator car.At the same time,the development of abnormal event monitoring in elevator car and early warning software is used to verify the stability and reliability of the algorithm in the actual environment,thereby establishing a high-performance elevator identification control system,and achieving the engineering of image restoration and object detection tasks.