Research On Low Light Industrial Image Enhancement Based On Neural Network

With the development of Internet of Things technology,factories nowadays generally use cameras instead of manual monitoring of equipment operation.However,the factory environment often has many characteristics such as uneven lighting,high levels of floating impurities in the air,and equipment shaking,resulting in blurred details and low quality of the final obtained images.In addition,in low illumination environments such as nighttime,the obtained images may have problems such as narrow dynamic range,missing detail information,and accompanied by a large amount of global noise.This type of image not only affects human visual perception ability,but also poses difficulties for computers to further process images.Therefore,aiming at the problem of low quality and low brightness of industrial scene images,this paper adopts infrared feature fusion,transfer learning and attention mechanism to study industrial image enhancement methods.The main research work of this article is introduced as follows:(1)In order to improve the image quality of low light images in the self-made dataset in this article,an industrial low light image enhancement algorithm using infrared feature fusion is proposed.Based on image fusion technology,visible light images and infrared images under low illumination are fused to improve image quality.Firstly,the saliency map of the infrared image is extracted and registered using the proposed SUSIFT algorithm.Then,the dual complex wavelet transform is used to decompose and fuse the infrared and visible light images in the frequency domain,achieving image enhancement of low-quality industrial images.(2)In order to ensure that the quality of industrial low light image can still be improved when the data set is small,Retinex Net network model is improved based on transfer learning theory,and Retinex Net TL network is proposed.And realized transfer learning of Retinex Net TL network from LOL dataset to self-made dataset in this paper.The main network framework,image decomposition module,and reflection and illumination component module were elaborated.Finally,subjective and objective tests were conducted on the algorithm’s effectiveness,and the test results showed its effectiveness.(3)In order to further improve the performance of the algorithm in small-scale data sets,the improved Retinex Net industrial low light image enhancement algorithm based on transfer learning is improved by introducing the attention mechanism module.Introduce the CBAM attention mechanism into the illumination component enhancement module and the reflection component denoising module,and design the overall framework model of the Retinex-Net-AM network.Through comparative experiments,the results showed that the PSNR index of the model increased by an average of 7.21%,and the SSIM index increased by an average of18.15%,verifying the effectiveness of the improved model.