高级检索

融合多级特征增强与权重网格统计的煤矿井下图像匹配

Coalmine image matching by fusing multi-level feature enhancement and weighted grid statistics

  • 摘要: 图像特征提取与匹配是实现煤矿井下视频和图像拼接、视觉定位导航等任务的一项关键技术,但受井下低光、不均匀光照以及重复纹理等环境因素的影响,相机采集到的图像往往对比度低、纹理信息不明显,存在特征点提取困难、误匹配率高的问题。对此,文章提出一种融合多级特征增强与权重网格统计的方法,以实现煤矿井下图像的有效匹配。首先,设计多级特征检测网络,在堆叠网络层的基础上引入可变形卷积层,保证特征的旋转不变性;其次,利用特征增强模块,将提取的特征点与描述子信息编码投影到高维空间,并通过Transformer网络增强特征间的可区分性;最后,采用基于权重网格的多阶段匹配优化策略,结合匹配质量因子与运动平滑性约束对初步匹配结果进行筛选,解决井下重复纹理感知混淆的问题,提高相似区域间的误匹配判识能力。在井下实际采集数据集和LOL、HPatches公开数据集上的大量试验表明:所提图像匹配方法具有更高的精度与鲁棒性。具体地,相较于ORB、SIFT、ASLFeat和Superpoint算法,所提特征提取方法的平均精度分别提升了33.07%、69.78%、17.65%和33.52%;相较于FLANN、BF+KNN、BF+RANSAC和BF+GMS特征匹配方法,所提特征匹配算法的平均精度分别提升了19.66%、23.26%、4.16%和18.46%。

     

    Abstract: Image feature extraction and matching is a critical technique for video and image stitching, visual localization and navigation in underground coal mines. However, underground tunnelling environmental factors, such as low light, uneven illumination, and repetitive textures, result in low-contrast images with indistinct texture information, making feature point extraction difficult and increasing the rate of mismatches. To address these issues, an coalmine image matching method based on multi-level feature fusion enhancement and weighted grid statistics is proposed. Firstly, multi-level feature detection network based on deformable convolution layers and stacked network is adopted for ensuring the rotational invariance of features. Secondly, both the extracted feature points and descriptors are encoded and projected into a high-dimensional space using a feature enhancement module, and the Transformer is then used to enhance the distinguishability of the features. Finally, to address the perceptual confusion problem caused by repetitive textures in underground scenes, a multi-stage matching optimization strategy based on weighted grids is used. This strategy can combine matching quality factors and motion smoothness constraints to filter the mismatches. A lot of experiments were conducted on an underground actual dataset, as well as on two public datasets including LOL and HPatches. The experimental results show that the proposed feature extraction and matching method has higher accuracy and robustness. Specifically, compared with the ORB, SIFT, ASLFeat and Superpoint algorithms, the proposed feature extraction method achieved average accuracy improvements of 33.07%, 69.78%, 17.65% and 33.52% respectively. Compared with the feature matching methods including FLANN, BF+KNN, BF+RANSAC and BF+GMS, the proposed feature matching method achieved average accuracy improvements of 19.66%, 23.26%, 4.16% and 18.46%.

     

/

返回文章
返回