An Image Edge Detection Algorithm Based on Multi-Feature Fusion

Edge detection is one of the core steps image processing and computer vision. Accurate fine edge will make further target semantic segmentation more effective. Holistically-Nested (HED) network has been proved to be a deep-learning with better performance for detection. However, it found that when HED used in overlapping complex multi-edge scenarios automatic object identification. There detected incomplete, not smooth other problems. To solve these problems, an algorithm based on improved feature fusion proposed. On hand, features are extracted using network: convolution layer improved. The residual variable block replace normal enhancement model extract from edges different sizes shapes. Meanwhile, empty original pooling expand receptive field retain global information obtain comprehensive information. Otsu algorithm: Otsu-Canny adaptively adjust threshold value scene achieve under optimal value. Finally, by fused final edge. Experimental results show Berkeley University Data Set (BSDS500) data set size (ODS) F-measure proposed 0.793; average precision (AP) 0.849; speed can reach than 25 frames per second (FPS), which confirms effectiveness method.