Edge Detection with multi-scale representation and refined Network

Edge detection is a representation of boundaries between objects and regions in an image. Due to the variations of types, scales, intensities as well as background, the detection of these boundaries represents a challenge for different computer vision algorithms. The exploitation of multi-scale representation using various deep learning backbones make edge detection more realistic. However, edge detection is still not effective in terms of erroneous detection in some scenarios and the quality of detected edges. In this paper, we attempted to overcome these challenges using a high-resolution and refined network by fusing the output of different scales of the network and the output of the first layers. The proposed network used horizontal and vertical blocks, while the horizontal block (HB) contains a set of convolution layers and the vertical block (VB) represent the different scale of the image using convolution and pooling layers. The output of each scale in VR is combined with the corresponding layer in HB. Then, we used an affined batch normalization layer as an erosion operation for the homogeneous region in the image before fusing these outputs for the final edge detection results. The proposed method is evaluated using the most challenging datasets including BSDS500, NYUD, and Multicue. The obtained results outperform the designed edge detection networks in terms of performance metrics and quality of output images.