xncaffe / caffe_penlu

Parameter Index Non-Linear Unit. It can alleviate the defect of neuronal death caused by the right side of the ReLU series function.

caffe_penlu

**This is the code implementation we implemented about the published paper "Identification of navel orange lesions by nonlinear deep learning algorithm".

**The download address of the paper is: https://www.scielo.br/j/eagri/a/vJBbggfFXQw7RkcLQMXN87D/?format=pdf&lang=en

We implement it under the caffe framework. For detailed caffe information, please refer to its official website： https://github.com/BVLC/caffe

How to use our code?

1. Download the official caffe package
2. Download our PEBLU package
3. Refer to PENLU's caffe.proto to modify the official caffe.proto (it is located under caffe/src/proto/)
4. Merge the include and src in PENLU to the official caffe
5. Compile caffe according to the official caffe compilation method, Please refer to: Caffe | Installation (berkeleyvision.org)
6. We provide some examples of our use of PENLU in network structures. For detailed network structure, please refer to caffe_penlu/models.
7. The network structure used by navel orange recognition is located in caffe_penlu/navel_orange/
8. The image data used for navel orange identification is stored in CSDN, click the link to download: https://download.csdn.net/download/xunan003/87283623?spm=1001.2014.3001.5503

Regarding the 3rd method of use, you need to add the registration part of the penlu layer to caffe.proto in caffe, as follows:

message LayerParameter {
  optional string name = 1; // the layer name
  optional string type = 2; // the layer type
  repeated string bottom = 3; // the name of each bottom blob
  repeated string top = 4; // the name of each top blob
  .......
  .......
  .......
  .......
  optional PENLUParameter penlu_param = 151; // 151 is the id, which is not the same as the id of other registration parameters
}
...
...
...
message V1LayerParameter {
  repeated string bottom = 2;
  repeated string top = 3;
  optional string name = 4;
  .......
  .......
  .......
  .......
  optional PENLUParameter penlu_param = 50; // 50 is the id, which is not the same as the id of other registration parameters
}
...
...
...
message PENLUParameter {
  optional FillerParameter alpha_filler = 1; // default = 1
  optional FillerParameter beta_filler = 2; // default = 1
  optional FillerParameter eta_filler = 3; // default = 1

  // Whether or not slope paramters are shared across channels.
  optional bool channel_shared = 4 [default = false];
}