Research On Low Light Image Enhancement Algorithm For Wellbores

With the introduction of the "carbon neutral" concept,smart mines are increasingly being used.Coal mining technology is gradually transforming from dangerous manual production to safe and intelligent production;from mechanized and automated mining to intelligent and unmanned mining.In this process,image processing technology plays an increasingly important role.Since the underground environment is characterized by uneven illumination,low brightness,mining dust,and electromagnetic interference,this leads to problems such as low contrast and high noise in the images captured in mines.These characteristics seriously hinder the application of intelligent detection and intelligent perception technology with image processing as the core.For the image enhancement of uneven illumination of the shaft wall,the following work is done in this paper: the acquisition and preprocessing of the shaft wall images and the production of the shaft wall image dataset(Wall dataset);the multi-branch fusion attention networks MANet and CC-UNet are proposed for the low-light image characteristics,respectively;the network performance evaluation and analysis are performed on the standard dataset and the Wall dataset.The low-light images are acquired through a low-light image acquisition platform on the wellbore wall.The acquired low-light images have motion blur and cannot be directly used for network training.The acquired images need to be corrected for deblurring and pillow distortion;then the images are cropped to obtain a standard image in the format of 256*256;finally,the cropped images are subjected to data enhancement operations to expand the dataset and obtain the borehole wall dataset.The multi-branch fusion attention enhancement network MANet is proposed for low-light images with uneven light distribution and easy overexposure.The network uses the U-Net network and combines depth-separable convolution with an attention mechanism to enhance low-light images.Compared to NEM,MANet has an 8.1%improvement in PSNR on the SICE dataset;BRISQUE has a better performance on the Well Wall Wall dataset with a 29.4% improvement,respectively,but still a significant improvement.Therefore,unlike MANet’s direct depth network approach,the CC-UNet network transforms the input image onto the HSV color space and makes the following optimizations based on the low-light image characteristics of the wellbore wall: CCAM improves on the large number of CBAM attention parameters,the tendency for the enhancement effect to be globally exposed,and the inability for channel and spatial information to flow.The use of MSR optimizes the depth of the separable Separate convolution is used to optimize the problem of extracting less effective feature information from well-wall images,and the saturation(S)channel is adjusted using adaptive gamma correction.Compared to MANet,the improvements in PSNR and SSIM are 7% and 1.8%,respectively;on the well wall Wall dataset,the improvements in PSNR and SSIM are 5.4% and 17.1%,respectively,compared to the TBEFN network,indicating that the CC-UNet network is more effective in low-light enhancement of well walls.Figure [40] Table [7] Reference [82]...