Machine Learning Notes Edge Detection Based on Deep Learning

一、Holistically-Nested Edge Detection

        Edge detection is a very commonly used technology in visual work. Traditional edge detection has included many classics such as Canny, Robert, etc., each of which has its own advantages. However, there is a problem that many parameters need to be manually adjusted. Therefore, deep learning research The researchers proposed an edge detection algorithm based on convolutional neural network.

        Namely HED, the algorithm performs image-to-image prediction by utilizing a deep learning model of a fully convolutional neural network and a deep supervised network. HED automatically learns rich hierarchical representations, which are important for resolving challenging ambiguities in edge and object boundary detection. Significantly improves the state-of-the-art on the BSD500 dataset (ODS F-score of .790) and the NYU Deep dataset (ODS F-score of .446), and improves speed (0.4 seconds per image).

The HED network removes the last 5 layers on the basis of the VGG network. The fully connected layer and the softmax layer behind the vgg are mainly used for classification. The HED network only needs to extract the features of the picture, and the previous convolutional layer and pooling layer are retained (note : remove the last pooling layer). And the feature map is merged and then 1*1 convolution, and finally activated by sigmoid.

        Paper address