Hed_ depth learning of edge detection

From the thesis, holistically-nested Edge detection, Iccv2015,marr Award nomination, very worth reading one.

The work of edge detection is divided into the following 3 research directions:

(1) Traditional detection operator: Sobel, Canny

(2) Manual features based on information theory: Statistical Edges, Pb, GPB

(3) manual features designed based on learning methods: BEL, Multi-scale, Sketch tokens, structured Edges

As can be seen from the above picture, the canny detection will exist disconnection line situation, and hed basically do not have this phenomenon, hed effect is obviously better than canny.

The whole nesting edge detection holistically-nested edge detection (HED).

Insight

(1) End to end: Image-to-image

(2) based on the improvement of FCN and VGG, 6 loss are optimized for training, the edges of different scale are exported through multiple side output, then the final edge output is obtained through a training weight fusion function. Can solve edge and object boundaries ambiguity

(3) Sample Unbalance treatment method: Class-balanced_sigmoid_cross_entropy

Network Overall structure:

The author in Conv1_2, Conv2_2, Conv3_3, Conv4_3,conv5_3 behind, and then access to Sigmoid_cross_entropy_loss, and at the last level, the above 5 levels of output did concat, At the same time, access to Sigmoid_cross_entropy_loss, so that all the loss are equal to the probability of training, so that the end to get a better model.

One thing to note here is that 6 loss are trained at the same time, and if you only train the last loss, the precision will drop.

In the prediction phase, the final result can be obtained directly from the output of the last layer. You can also get the output of all the layers and then find an average as the final result. The advantage of this is that the precision will be further improved, the disadvantage is to add additional operations, time will increase.

As shown in the following figure, the side-output1-5 is the output of the 第1-5个 convolution layer respectively. Fusion-output is the last layer of output, average1-4 for the output of the 第1-4个 convolution, average1-5,average2-4,average2-5, and so on. Merged result for the results of all layers to find the results of the average merge.

can see that

(1) For each individual layer, the final layer of the feature fusion results are the best.

(2) to average the results of all layers, the final accuracy rate will be improved.

During the training process, the edge detection is actually a 2 classification task for each pixel. Most of the pixels are not edges, only a few pixels are marginal, and in order to balance the positive and negative samples, the authors introduce class-balanced_sigmoid_cross_entropy.

The figure above is the formula in the author's paper, where the author does not write the value of the label Y_hat. be organized as follows.

Sigmoid_cross_entropy:

-y_hat* log (sigmoid (y))-(1-y_hat) * log (1-sigmoid (y))

Class_balanced_sigmoid_cross_entropy:

-β*y_hat* log (sigmoid (y))-(1-β) * (1-y_hat) * log (1-sigmoid (y))

of which, | Y-| as the edge of the label,| Y+| is not an edge label

β= | y-|/| y|

1-β= | y+|/| Y |

Finally, the paper realizes the Depth results of the BSD500 dataset (ODS F-score of:782), NYU f-score DataSet (ODS of:746 gpu:0.4s,cpu:12s), State-of-the-art.

References:

Https://github.com/s9xie/hed