Classify
Information
DropoutLayer
ConvolutionLayer
BatchNormalizationLayer
ImageAugmentationLayer
AggregationLayer
LinearLayer
SoftmaxLayer
NetEncoder
NetDecoder
NetInitialize
NetTrain

Image Processing
Machine Learning

Distinguish Daytime from Nighttime Pictures

Train a simple classifier to distinguish daytime photos from nighttime ones

Open in Cloud
Download Notebook

Collect photos taken during daytime and nighttime:

In[1]:=

training=ResourceData["Example""Distinguish Daytime from Nighttime Pictures","TrainingData"]

Out[1]=

Use

Classify

to automatically train a classifier function to distinguish between them:

In[2]:=

daynight=Classify[training]

Out[2]=

Useful data and properties about the classifier function are available via

Information

In[3]:=

Information[daynight]

Out[3]=

Consider a new list of pictures:

In[4]:=

test=

;

Classify the new set of data:

In[5]:=

daynight[test]

Out[5]=

{Day,Night,Day,Night,Night,Night}

Analyze the probabilities of a misclassified example:

In[6]:=

daynight

,"Probabilities"

Out[6]=

Day1.19968×

-9

,Night1.

Alternatively, manually define a convolutional network with minimal data augmentation and regularization (dropout, batch normalization) to perform the classification:

In[7]:=

convBlock[n_]:={DropoutLayer[],ConvolutionLayer[n,3,"Stride"2,PaddingSize1],BatchNormalizationLayer[],Ramp};

In[8]:=

daynightNet=NetChain[{ImageAugmentationLayer[{120,120},"ReflectionProbabilities"{0,.5}],convBlock[16],convBlock[32],convBlock[64],convBlock[128],AggregationLayer[Max],LinearLayer[2],SoftmaxLayer[]},"Input"NetEncoder["Image"],"Output"NetDecoder[{"Class",{"Day","Night"}}]]

Out[8]=

NetChain

uniniti

aliz

Inputport:	image
Outputport:	class



Note that encoders and decoders are used to convert input and output to and from numerical arrays. This way, the network works like a normal function:

In[9]:=