摘要
针对较为复杂的光伏电站场景中的图像中光伏板边缘提取问题,本文提出了一种精细检测方法。所提出的方法首先引入多尺度的图像特征,通过鼓励网络不同层输出的特征表示检测对应该层尺度的图像边缘,进而充分融合不同尺度特征中所包含的丰富信息,使得对于不同尺度物体边缘的检测都能够更加精细;其次,本方法利用多任务学习结构,通过挖掘语义分割和边缘检测两个任务之间的相关性与互补性,使得图像特征更加聚焦于待提取边缘的物体,排除场景中无关物体或噪声的干扰,从而生成更加合理的图像边缘。量化的实验结果与可视化结果均表明,本方法能够获得更加精准的图像边缘检测效果。
Aimed at edge extraction of photovoltaic panels in the complicated photovoltaic power stations, this paper puts forward a precise detection method. For the proposed method, first, multi-dimensional image features are introduced. The feature representations from different layers of network detect image edge correspond to the dimension of this layer. The plenty of information in different dimension features is fully integrated to make the edge detection of object with different dimensions be more precise. Second, this method makes use of multi-task learning structure and digs out the correlation and complementarity between semantic segmentation and edge detection to make low-level features of image more concentrate on the object edge to be extracted and eliminate the interference of irrelevant objects or noise in the scene, thereby generating a more ideal image edge. Both quantitative experimental result and visualization result show that this method can acquire more precise image edge detection effect.
出处
《计算机科学与应用》
2021年第11期2599-2608,共10页
Computer Science and Application
