jonwomack/holographic-plankton-classification

holographic_plankton_classification

Automated Plankton Classification from Holographic Imagery with Deep Convolutional Neural Networks Citation If you find this work or code is helpful in your research, please cite:

title={Automated Plankton Classification from Holographic Imagery with Deep Convolutional Neural Networks}, author={Buyu Guo and Lisa Nyman and Aditya R. Nayak and David Milmore and Malcolm McFarland and Michael S. Twardowski and James M. Sullivan and Jia Yu and Jiarong Hong},

Table of Contents

• Installation
• Instructions for Use
• How to use

Installation

The following Python libraries should be installed with pip:

• thorch
• torchvision
• matplotlib
• trackpy
• opencv
• sklearn

Example of how to install with pip:

$ pip3 install --user sklearn

Instructions for Use

##Training

1. Create training and validation datasets. These should be a txt file with columns for image ID and their correspoding classes.
2. Modify training parameters in [teShufflenet15.py.py]
3. Train the model

Parameters in [teShufflenet15.py]:

• batchSize: he number of training examples utilized in one iteration.

• numClasses: how many classes we have.

• numEpoch: the number of passes of the entire training dataset the machine learning algorithm has completed.

• testSize: the ratio to split your data into training and testing

• learningRate: a tuning parameter in an optimization algorithm that determines the step size at each iteration while moving toward a * * minimum of a loss function.

• dropNum: ignoring units (i.e. neurons) during the training phase of certain set of neurons which is chosen at random.

• featureExtract = True

• usePretrained = False

• dropout = True

• mono: image type is grayscale or RGB

• savePreAndRealFlag: a parameter for

• inputSize: size of training images. If their length/width is smaller the inputSize, it will be padded to inputSize

• imgType: type of the image

• txtFil: txt file generated in step 1

• modelName: the name of output model

• filePath: set the path

• savePath：set the saving path

##Classifying

1. Modify training parameters in [groupProcessing.py]
2. Processing

• holoPath: path of holograms

• netPath: path of the network

• classNum: how many classes we need to classify

• fileType: image type

• segmentSavePath: the save path for segments (cropped from holograms)

• cropSizeLimit: segments with length or width less than cropSizeLimit will not be considered

• morphFlag: whether to use morphological analysis

• adaptiveThreshold: whether to use adaptive threshold

• noSuperviewThre: whether to use morphological analysis

• outputRawImg: whether to output the original holograms

• outputCoordinateFlag: whether to output the coordinates of segments

• saveWithBackgroundFlag = True

• paddingNum: add several pixels around segments

Please refer to the opencv documentation for the following parameters：

• kernalSizeBlur: https://docs.opencv.org/master/d4/d13/tutorial_py_filtering.html

• adaptPare1: https://docs.opencv.org/master/d9/d61/tutorial_py_morphological_ops.html

• adaptPare2: https://docs.opencv.org/master/d9/d61/tutorial_py_morphological_ops.html

• morphPara1: https://docs.opencv.org/master/d9/d61/tutorial_py_morphological_ops.html

• outputPath: output path

• inputSize: no need to be same with training para

• batchSize:

• saveClsaaFlag: whether to save classes separatly

• printScoreFlag: whether to print the classification score

• paddingFlag: whether to pad the images

• topFlag: whether to save the top score

• respectivelyCountFlag: whether to save the class of segments separatly

How to use

• Train:

1. Mark species as numbers (e.g., C. debils is 0; Diatom sp. is 1; D. brightwelli is 2 etc.)
2. Create a txt file map the images path with their species (shown in the figure)
3. Set parameters in ‘teShufflenet15.py’
4. Run ‘teShufflenet15.py’

• Classify:

** Classify the full-size holograms

1. Set parameters in ‘groupProcessing.py’
2. Run ‘groupProcessing.py’

** Classify the segments

1. Comment out ‘imgCrop.imgSegment(…)’ in ‘groupProcessing.py’
2. Set parameters in ‘groupProcessing.py’
3. Run ‘groupProcessing.py’