DL_EXP_PC

Deep learning experiment - pixel classification

A simple example of the creation and application of a neural network to pixel classification using keras.

chapter	content	open in colab	open locally
Pixel classification	example of a small fully connected network that classifies pixel in images	DL01	DL01
01 - installation	keywords, functions, modules and packages, booleans, if-statement, help, shell-commands	Cell01	Cell01
02 - paths, patch and sample size	variables, integers, floats, strings	Cell02	Cell02
03 - reading the input and ground-truth image	loading and displaying tif-images	Cell03	Cell03
04 - getting foreground and background pixels	tupels, lists, numpy arrays and the python imaging library (PIL)	Cell04	Cell04
05 - random sampling of training data	pseudo random numbers, for-loops, while-loops, objects and classes, random sampling	Cell05	Cell05
06 - preparation of training data	creating the network input from intensities in pixel neighborhoods	Cell06	Cell06
07 - shuffle and deal	shuffling feature vectors and labels	Cell07	Cell07
08 - creating the network	keras, sequantial model, adding layers, compiling, activation, metrics, loss, optimizer, visualization	Cell08	Cell08
09 - training the network	fitting the model, history of loss and accuracy, validation split, epochs, batch size, testing	Cell09	Cell09
10 - visualizing performance	creating plots with matplotlib and interactive plots with mpld3	Cell10	Cell10
11 - classifing a single feature vector	predict, predict_classes, flatten	Cell11	Cell11
12 - extracting feature vectors from images	manually setting weights, saving and loding a model, neighborhoods to features	Cell12	Cell12
13 - creating the output folder and getting the input paths	isdir, listdir, join paths, list comprehension, sorting	Cell13	Cell13
14 - classifying pixels in images	image segmentation, thresholding, cropping, writing tiff images	Cell14	Cell14
15 - visualizing results	matplotlib, pyplot, subplots and imshow	Cell15	Cell15
A one unit network	an example with the smallest network possible containing only one unit	Bonus track	Bonus track