Research On Unsupervised Image Enhancement Algorithm Technology And Application In Complex Scene

Low-light image enhancement aims to capture and enhance visual perception of data through digital image processing techniques,in situations where lighting conditions are insufficient,in order to obtain more information and better aesthetic effects.However,in the field of low-light image enhancement,it is difficult to collect large-scale paired data samples because low-light and normal lighting images of the same scene are difficult to capture,making it difficult for deep learning to fully leverage its advantages due to the lack of effective supervision.Therefore,a series of unsupervised learning algorithms that do not require paired data have been proposed.However,existing unsupervised methods still suffer from insufficient authenticity and unclear enhancement effects for images under extremely dark conditions.To address these issues,this paper proposes an unsupervised low-light image enhancement method that uses an efficient cyclic generative adversarial network(GAN)to achieve low-light image enhancement.The proposed network includes a global-local generator and a two-stage discriminator.First,the generator is designed as a two-branch global-local generator,with the two branches merging to fully learn the global and local information of the image.Second,the discriminator is designed as an adaptive two-stage classifier that adapts to locate the local regions that need to be re-evaluated by discriminating the global regions,ultimately generating global and local discriminant masks.Finally,this paper uses a focus frequency domain loss function to enable the model to adaptively focus on difficult-to-synthesize frequency bands and avoid image distortion.The proposed method was extensively tested on six benchmark datasets,and the results demonstrate the superiority and effectiveness of this method.In response to the noise and color distortion issues in the algorithm of the previous work,this paper proposes an unsupervised low-light enhancement model framework based on histogram equalization prior.This framework integrates the physical model of image lighting conditions and residual learning through a shared-weight cascade learning low-light image enhancement module,enabling enhancement of low-light image scenes in different environments.A result calibration module is also designed to accelerate the convergence of the model,enabling it to converge to superior results faster and more effectively.Furthermore,this work proposes a generative adversarial denoising module that separates noise and content in the enhanced image by using unparsed clean images,resulting in cleaner enhanced images.Extensive experiments demonstrate the superiority of this work in terms of performance,surpassing most recent algorithms,including traditional,supervised,and unsupervised methods.Finally,this paper extends the work of the first algorithm to the practical application of medical endoscopic images,achieving simulation of indocyanine green staining in white light endoscopic images through global-local generative adversarial learning networks,enhancing the details of medical image lesions.This work uses a thumbnail instance normalization method to extract features from endoscopic images,dividing large-scale images into multiple small blocks and inputting them into a cyclic adversarial neural network,minimizing the consumption of computational resources and time while ensuring the consistency of image style.Experimental results and expert evaluations demonstrate that the simulation effect of indocyanine green staining in white light endoscopic images achieved in this work is superior to traditional indocyanine green staining,and can assist endoscopic doctors in the auxiliary treatment of different types of lesions during endoscopic examinations.