基于深度融合卷积神经网络的图像边缘检测

Image edge detection based on deeply⁃fused convolution neural network

图像边缘检测是数字图像分析领域的一项重要研究内容。受图像拍摄条件、图像内容自身复杂性、图像内容与背景接近程度等多种因素的影响,图像的边缘线检测容易发生漏检、误检。针对此问题,文中提出一种卷积神经网络结构算法,以提升图像边缘检测效果和质量。首先,对输入图像提取出五类不同层次水平、尺度的卷积特征;然后,按照相邻尺度将每三类卷积特征分成一组,通过逐步转置的方式依次尺寸对齐再融合;再对三组融合结果特征进行二次深度融合;最后,基于融合卷积特征并运用卷积操作实现边缘线检测,采用指标Optimal Dataset Scale(ODS)、Optimal Image Scale(OIS)、Average Precision(AP)度量图像边缘检测的质量。结果表明:在BSDS500数据集上,ODS、OIS、AP三个指标的得分分别为0.815,0.832,0.851;在NYUD数据集上,得分分别为0.7620,0.7700,0.7819。与其他同类算法相比,所提算法指标分值更高,能够提升图像边缘检测质量。

Image edge detection is an important research content in the field of digital image analysis.The edge line detection of images is prone to miss detection and false detection duo to the influence of image shooting conditions,complexity of image content,similar level of image content and background and other factors.In allusion to these questions,a convolution neural network structure algorithm is proposed to improve the effect and quality of image edge detection.Five kinds of convolution features with different levels and scales are extracted from the input image.Each three types of convolutional features are merged into a group according to the adjacent scale,which are aligned and refused by means of step⁃by⁃step transposition,and then the secondary depth fusion are performed according to the features of three groups fusion results.The edge line detection can be realized based on the fused convolution features and by means of the convolution operation.The indicators of optimal dataset scale(ODS),optimal image scale(OIS)and average precision(AP)are used to measure the quality of image edge detection.The results show that the scores of ODS,OIS,and AP are 0.815,0.832 and 0.851 respectively in the BSDS500 dataset,and 0.7620,0.7700 and 0.7819 respectively in the NYUD dataset.In comparison with the existing algorithms,the proposed algorithm has a higher indicator score and can improve the quality of image edge detection.