Research On Image Super-resolution Fusion Model Using Unsupervised Decoupled Generative Learning

Under the background of Industry 4.0,the intelligent automation industry is highly developed,in which most of the machine intelligence is realized based on vision perception systems.Therefore,images are gradually becoming the main carrier of information to guide system decision making and controlling,also making high quality image acquisition a prerequisite for making accurate decisions.At the same time,daytime and nighttime lighting conditions are the most common external environments for industrial imaging systems,and the most commonly-used image acquisition systems today are those based on visible and infrared wavelengths.For the daytime environment commonly used imaging systems in the visible band,which focus on visible light imaging at 380-760 nm wavelengths,can capture reflected light and ensure rich textural detail of the target.Due to limitations in cost,reliability and application scenarios,it is difficult to use optical zoom imaging devices with large depth of field in practical application scenarios.This makes it difficult to focus targets located in multiple focal planes simultaneously in a single image through a single imaging session.Multi-focus image fusion technology can extract the focus area from several different target focus images in the same scene and reconstruct it into a single full-focus image,providing a richer all-in-focus image input for downstream application scenarios.As for imaging systems in the infrared band,which is commonly used for nighttime environments,its focus on wavelengths between 8 and 14 μm can reflect the heat radiated by an object and highlight hot radiating objects even under conditions of insufficient illumination or severe occlusion.As a result,infrared and visible image fusion also have wider applications.However,for both of these image fusion techniques,low resolution due to the limitations of the imaging principle is the core factor affecting the quality of fusion,so image super-resolution fusion offers new ideas for fusion techniques.This paper focuses on the key techniques and issues involved in multi-focus image fusion algorithms and visible-infared image fusion algorithms,and proposes an innovative new framework for simultaneous implementation of image fusion and super-resolution based on unsupervised decoupling generation learning,and applies it to both fusion applications for validation.The main work is as follows:First,an overview of multi-focus and infrared image acquisition devices.The limitations of traditional methods are considered to design a multi-focus and infrared image acquisition scheme that is consistent with the methods in this paper.This paper focuses on the acquisition of multi-focus images to be fused,using optical field imaging and digital refocusing to acquire images to be fused.Considering that the refocusing algorithm is computationally large and very time-consuming to debug,this paper first designs and implements Open Refocus’ s QT-based light field parallel refocusing software to accelerate the processing speed of the space domain digital refocusing algorithm and simplify the processing of the algorithm parameters.Subsequently,this paper qualitatively and quantitatively evaluates the Open Refocus refocusing results on public datasets and real experimental scenarios,respectively,and can demonstrate that the proposed method is 2 times faster than the traditional model in CPU mode and 8 times faster in GPU mode with similar reconstruction quality.Second,based on the idea of deep image prior,this paper proposes an unsupervised decoupled generative learning framework for image super-resolution fusion,combining the coupling relationship between the parts of the physical model.The framework uses generative networks to generate each decoupling component of the physical model,and then designs the coupling loss function through the physical model to form a closed loop,and finally achieves super-resolution fusion with only one set of low-resolution input data.Third,in the task of multi-focus image super-resolution fusion,this paper first unifies the physical model of multi-focus image fusion and super-resolution.On this basis,this paper first proposes a semi-supervised learning based REPAID two-stage multi-focus image super-resolution fusion model,and the proposed method can achieve a resolution improvement of about 37μm compared with the traditional method.Subsequently,a focus segmentation decoupling deep fusion prior is proposed in an unsupervised decoupled generative learning framework to achieve high-quality multi-focus image super-resolution fusion simultaneously in only one stage.This paper also presents a qualitative and quantitative evaluation of the focus segmentation decoupling model.This paper also evaluates the focus segmentation decoupling model qualitatively and quantitatively,both of which are able to achieve better results compared to mainstream methods.Fourth,in the task of super-resolution fusion of visible-IR images.This paper proposes a Retinex-based deep adaptive super-resolution fusion method based on unsupervised decoupled generative learning,which simultaneously achieves high-quality and adaptive super-resolution fusion of visible-IR images using only one model.Finally,the proposed model achieves an optimal structure in all qualitative comparisons and quantitative metrics,with quantitative metrics outperforming existing mainstream methods by approximately11.76%.