Study On Low Illumination Image Enhancement Of Field Relics Based On GAN

Image data is widely used as an important medium for storing information in areas such as computer vision,security surveillance and medical examinations.However,information recorded from low-light images collected at night or in poorly lit environments can be seriously lost,affecting the use of the data.Particularly in the field of field heritage security,these images taken at night are extremely poorly visualised and prone to problems such as insufficient brightness,high noise levels and blurred detail sections,preventing the direct acquisition of clear and usable images,to the detriment of heritage security.In order to improve the quality of visual perception of data in poorly lit scenes to obtain more information,low-light image enhancement techniques have emerged.This paper therefore investigates low-illumination image enhancement techniques for field artefacts,aiming to lay the foundation for higher-level image vision tasks by investigating algorithms to enhance image brightness,recover image colour and detail information,and improve generalisation capabilities.The main work of this paper is as follows:(1)Based on the existing Enlighten GAN model,a generative adversarial network model ULEGAN based on the attention mechanism is proposed to recover the brightness and details of images.The generative network part is improved by using a U-Net++network structure capable of extracting deep features along with shallow features,enabling more adequate information transfer between feature maps of different sizes and improving the performance of the model.In addition,a new calculation method for the luminance attention mechanism is used to better guide the model to enhance dark regions and avoid over-enhancing bright regions.By designing different comparison experiments to compare the algorithm in this chapter with several commonly used deep learning-based enhancement algorithms,good results are achieved in both human subjective vision and objective evaluation metrics,and the recovered images are improved in both luminance and content recovery compared to the Enlighten GAN algorithm.(2)Further improvements were made to ULEGAN to address the problems that exist in it.First,the generative network part uses the U-Net3+ network structure,which reduces the number of nodes added in the middle and simplifies the model structure.Also,having each decoder layer fuse small-scale feature maps from the same scale,large scale and encoder allows the network structure to capture full-scale feature map information and explore enough information.Secondly,the discriminator network part uses a multidiscriminator network structure instead of the original discriminator network,which is able to solve the problem of poor image colour and texture recovery by examining image features at different scales.The multi-discriminator network structure consists of a colour discriminator,a texture discriminator and a multi-scale discriminator,where the colour discriminator focuses on the colour part of the image to ensure that the colour of the generated image is closer to the real image,and the texture discriminator focuses on the sharpness of the edges of the generated image.Finally,the loss function section uses a combined loss function to allow the model to capture more of the image content in order to obtain better low-light enhancement.The experiments show that the algorithm in this chapter achieves the best results in terms of visual perception,and the enhanced image brightness,colour recovery and detail information are recovered to the maximum extent,and the objective evaluation index is also better than other algorithms.