低照度隧道口视觉融合技术研究

Visual Fusion Technology of Dim-lightening Tunnel Entrance

针对隧道出入口光照强度不足导致成像质量不佳的问题,提出了一种基于低照度隧道口红外与可见光图像融合方法。首先利用双边滤波与光照分量,对隧道口低照度红外和可见光源图像进行自适应图像增强;其次通过非下采样轮廓波进行多尺度、多方向分解以弥补预处理后的图像信息损失;在低频系数上,采用基于卷积稀疏表示与局部能量特征相结合的方法进行融合;在高频系数上,根据底层视觉特征构建新活性度量方法与光谱边缘处理;最后,将得到的低频和高频融合层进行重构得到最终的融合图像。实验结果表明,所提出的融合算法与BF、SE、NSCT-BF、SF-Energy-Q、SR-C&L五种算法相比,主观上视觉效果更好,辨识度高,整幅图像场景得以凸显,互信息量、信息熵、标准差均为最高,分别为7.5962、7.7642、82.1941,运算时间至多减少0.0232 s。该方法在降低噪声、均衡光照、恢复细节方面有参考意义。

In highway tunnels in mountainous areas,there is insufficient illumination intensity in a closed environment at night,and after imaging,the average pixel illumination intensity is low.The data information obtained by a single sensor is usually limited.Multiple sensors improve the image fusion performance at the tunnel mouth with low illumination.Infrared sensors use the thermal radiation generated by the object to achieve automatic detection and capture the object under the condition of low illumination;visible images provide rich background information.The image information of infrared and visible light and the electromagnetic spectrum is fused to obtain enhanced and more comprehensive scene information.Image processing in a low-illumination environment has always been a hot issue in academic research.This paper used Convolution Sparse Representation(CSR),Spectral Edge(SE)and local energy features for image fusion.An intelligent sensing method for spatial information of highway tunnels under static and dynamic light environments is proposed.The denoising and fusion are processed simultaneously to avoid the loss of visible and near-infrared information during fusion processing.Bilateral filtering and light component are used for adaptive image enhancement of low-illuminance infrared and visible light source images at the tunnel mouth.Gamma correction is used to correct the illumination component to avoid distortion during image enhancement.In order to improve the visual information presented by visible tunnel light,infrared and original visible image are fused to enhance the dark details of infrared pixels.In order to further improve the feedback of multiple information in the image,the non-subsampled contour is used to decompose the preprocessed image in multi-scale and multi-direction.The non-subsampled pyramid and non-subsampled directional filter are the main components of the non-subsampled contour wave.The klayer decomposition of the preprocessed source image,k+1 subband image with the same size can be obtained.The algorithm uses bilateral filters to decompose a single low-frequency subgraph decomposed by k layers into low-frequency basic components and detail feature components,respectively,for visible image and near-infrared images.The former is fused by local energy features,while the detail feature components are fused by convolution sparse representation strategy.The weighted local energy preserves structured information.Since simple weighting often leads to fading of infrared targets,the local feature energy ratio is used to measure the details extracted to maintain the brightness of fusion targets.A new activity measurement method and spectral edge processing were constructed at the high-frequency coefficients according to the underlying visual features;edge information is injected into the multi-source image to extract high-frequency information.Finally,the fusion coefficients were reconstructed to obtain the fused image.Four groups of visible and infrared source images captured by simulating the driver’s line of sight were fused and compared with the algorithm results.The experiments were compared and analyzed from subjective evaluation and objective evaluation.Experimental results show that the CSR-SE-Energy algorithm overcomes the traditional"SR"and"pseudo-Gibbs"effects,makes up for the shortcomings of poor correlation between images,and saves Energy information and edge details.The fusion algorithm outperforms BF,SE,NSCT-BF,SF-Energy-Q and SR-C&L in subjective evaluation.The subjective visual effect has high contrast and good identification,the whole image scene can be highlighted,and the running time can be shortened.In objective evaluation,the highest MI value was 7.5962,the highest IE value was 7.7642,and the highest standard deviation value was 82.1941.Compared with BF,SE,NSCT-BF,SF-energy-Q and SR-C&L algorithms.This method has significant reference significance in reducing noise,equalizing illumination and restoring details.When processing the image at the entrance and exit of the low illumination tunnel,the operation time is reduced by 0.0232 s at most,reducing the overall operation cost and improving the image’s robustness and visual clarity.