Research On Target Recognition Algorithm In The Tunnel Dark Environment Based On Image Enhancement

Tunnel defect detection is an important part of tunnel maintenance,but due to the low illumination in tunnels,the captured images have problems such as low brightness and inconspicuous detail information.The existing low-light image enhancement algorithms often suffer from color distortion and loss of small details,resulting in poor overall visual effects that are not conducive to subsequent processing tasks.In this paper,two enhancement algorithms are designed for tunnel low-illumination captured images,and a set of tunnel complex dark environment defect detection schemes is designed as follows:1.A U-Net network-based low-light image enhancement algorithm in HSV space is designed,firstly converting the image from RGB space to HSV space,weakening the strong correlation between image color information and luminance information,and then separately enhancing the luminance component V to avoid color distortion in the enhanced image;using the structural feature of the improved U-Net network with same-layer feature jump connection,extracting the deeper part of the image The deeper features of the luminance component V are extracted by using the improved U-Net network with same-layer feature hopping connection to avoid the problem of feature loss due to forward propagation and gradient update of the algorithm.enhanced image.The objective indexes of the enhanced image are AG: 6.815,EN: 2.065,PSNR: 10.526,and SSIM: 0.719.2.A U-Net network-based multi-branch cyclic generative adversarial network lowillumination enhancement algorithm is designed,using a cyclic generative adversarial network for unsupervised training to enhance images without pairing low-quality and highquality image datasets,which can effectively avoid the problem of difficult image acquisition in dark tunnel environments.The multi-branch generator structure is designed for input image feature extraction,and the improved U-Net network is used as the main branch of the generator to solve the problem of limited image brightness enhancement,and the channel attention module ECA model added to it can effectively enhance the image deep detail information,while the residual convolutional network model is designed as the auxiliary branch of the generator to extract the image shallow feature information;a combination of global plus local The discriminator model is used to fully discriminate the "authenticity" of the image features;after the generator and discriminator continuously learn against each other,the final generator can produce high-quality images with good visual effects.3.To verify the practicality of the enhancement algorithm,a tunnel dark environment defect detection system based on image enhancement is designed.Firstly,the original low illumination images collected from the tunnel are enhanced to improve the image quality and readability,and to produce a high-quality target recognition dataset to ensure that the subsequent defect detection task can be carried out effectively;then the Faster R-CNN algorithm is designed as the main model of the detection system,and the recognition training is carried out through the dataset obtained in the early stage to obtain a high accuracy defect detection system that meets the task requirements;finally,an Finally,an interactive interface is designed to visually demonstrate the detection effect of the algorithm on tunnel defects.