Optimal and efficient machine learning method for deep semantic segmentation
View Patent ↗Four computerized machine learning methods for deep semantic segmentation are fast machine learning method, active machine learning method, optimal machine learning method, and optimal transfer learning method. The fast machine learning method performs a fast deep semantic segmentation learning on training images to generate a deep model. The active machine learning method performs a fast deep semantic segmentation learning on initial training images to generate a first deep model and then an active deep semantic segmentation learning to generate a second deep model. The optimal machine learning method performs a fast deep semantic segmentation learning on initial training images to generate a first deep model and then an optimal deep semantic segmentation learning to generate a second deep model. The optimal transfer learning method applies a pre-trained first deep model on transfer training images and then an optimal deep semantic segmentation transfer learning to generate a second deep model.
1. A computerized active machine learning method for deep semantic segmentation, comprising the steps of:
a) inputting a plurality of initial training images and corresponding initial partial ground truth data into electronic storage means;
b) performing a fast deep semantic segmentation learning- 1 by electronic computing means using the plurality of initial training images and the corresponding initial partial ground truth data to generate a deep model- 1 ;
c) inputting at least one test image into electronic storage means;
d) performing an active deep semantic segmentation learning by electronic computing means using the plurality of initial training images, the corresponding initial partial ground truth data and the test image to generate a deep model- 2 ;
wherein the active deep semantic segmentation learning comprises the steps of:
applying the deep model- 1 to the test image to generate a confidence map;
identifying and cropping at least one low confidence region image from the test image;
acquiring low confidence region new ground truth data from additional truth labeling; and
performing a fast deep semantic segmentation learning- 2 by electronic computing means additionally using the low confidence region image and the low confidence region new ground truth data to generate the deep model- 2 .
2. The computerized active machine learning method of claim 1 , wherein the plurality of initial partial ground truth data contains do-not-care regions.
3. The computerized active machine learning method of claim 1 , wherein the deep model- 1 and the deep model- 2 are encoder-decoder networks.
4. The computerized active machine learning method of claim 1 , wherein the fast deep semantic segmentation learning- 1 minimizes a do-not-care robust loss function at an output layer by iteratively adjusting parameters at each layer of the deep model- 1 .
5. The computerized active machine learning method of claim 4 , wherein the parameters are adjusted by a back propagation method.
6. The computerized active machine learning method of claim 1 , wherein the fast deep semantic segmentation learning- 2 performs a transfer learning using parameters from the deep model- 1 .
7. The computerized active machine learning method of claim 1 , wherein the fast deep semantic segmentation learning- 2 minimizes a do-not-care robust loss function at an output layer by iteratively adjusting parameters at each layer of the deep model- 2 .
8. The computerized active machine learning method of claim 7 , wherein the parameters are adjusted by a back propagation method.