Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision
Previous revision
Last revision Both sides next revision
cryolo_nets [2018/12/18 15:16]
twagner [Introduction]
cryolo_nets [2018/12/24 10:15]
twagner [Introduction]
Line 6: Line 6:
  
 The main components are **convolutional operations** and **max pooling operations**:​ The main components are **convolutional operations** and **max pooling operations**:​
-  * Convolutional layer: A convolutional learn local patterns. For 2D images, these are patterns in small region ​in of the input image. In the YOLO architecture,​ these are 1x1 or 3x3 windows. The output of a convolutional layer is called feature map.+  * Convolutional layer: A convolutional learn local patterns. For 2D images, these are patterns in small region of the input image. In the YOLO architecture,​ these are 1x1 or 3x3 windows. The output of a convolutional layer is called feature map.
   * Max-pooling operation: Max-pooling operations downsampling the feature maps of previous layers. This enables the following convolutional layers to see a larger extends of the input image.   * Max-pooling operation: Max-pooling operations downsampling the feature maps of previous layers. This enables the following convolutional layers to see a larger extends of the input image.
 Another characteristic of this architecture is the passthrough connection between 13 and 21. it helps the network to utilize low level features during detection. Another characteristic of this architecture is the passthrough connection between 13 and 21. it helps the network to utilize low level features during detection.
Line 107: Line 107:
  
 ==== Network #3 PhosaurusNet ==== ==== Network #3 PhosaurusNet ====
-At some point, we realized that the patch mode introduced a problem on images ​on with particles on carbon ​edges: CrYOLO picked sometimes particles on the carbon edge. As with patch mode, the network did not see the complete micrograph anymore, which made it harder to reliably detect the carbon edge. Moreover, passing multiple patches through the network is computationally more expensive as processing whole micrograph in a single pass. We had another idea to circumvent this problem which combines the advantages of the crYOLO and the YOLO networks. PhosaurusNet upscales the feature map after layer 21 by a factor of two and then concatenates the feature map of layer 13. This network finally picked small particles, generalized well for our training data, does not pick on the carbon edge, still sees the complete micrograph and is faster.+At some point, we realized that the patch mode introduced a problem on images with particles on the carbon ​edge: CrYOLO picked sometimes particles on the carbon edge. As with patch mode, the network did not see the complete micrograph anymore, which made it harder to reliably detect the carbon edge. Moreover, passing multiple patches through the network is computationally more expensive as processing whole micrograph in a single pass. We had another idea to circumvent this problem which combines the advantages of the crYOLO and the YOLO networks. PhosaurusNet upscales the feature map after layer 21 by a factor of two and then concatenates the feature map of layer 13. This network finally picked small particles, generalized well for our training data, does not pick on the carbon edge, still sees the complete micrograph and is faster.
 {{ :​phosaurus.png?​600 |}} {{ :​phosaurus.png?​600 |}}
 We recommend to use this network for single particle picking. For picking filaments, YOLO in patch mode still seems to work better. We recommend to use this network for single particle picking. For picking filaments, YOLO in patch mode still seems to work better.