Main Content

Preprocess Data for Domain-Specific Deep Learning Applications

Data preprocessing is used for training, validation, and inference. Preprocessing consists of a series of deterministic operations that normalize or enhance desired data features. For example, you can normalize data to a fixed range or rescale data to the size required by the network input layer.

Preprocessing can occur at two stages in the deep learning workflow.

  • Commonly, preprocessing occurs as a separate step that you complete before preparing the data to be fed to the network. You load your original data, apply the preprocessing operations, then save the result to disk. The advantage of this approach is that the preprocessing overhead is only required once, then the preprocessed images are readily available as a starting place for all future trials of training a network.

  • If you load your data into a datastore, then you can also apply preprocessing during training by using the transform and combine functions. For more information, see Datastores for Deep Learning. The transformed images are not stored in memory. This approach is convenient to avoid writing a second copy of training data to disk if your preprocessing operations are not computationally expensive and do not noticeably impact the speed of training the network.

Data augmentation consists of randomized operations that are applied to the training data while the network is training. Augmentation increases the effective amount of training data and helps to make the network invariant to common distortion in the data. For example, you can add artificial noise to training data so that the network is invariant to noise.

To augment training data, start by loading your data into a datastore. For more information, see Datastores for Deep Learning. Some built-in datastores apply a specific and limited set of augmentation to data for specific applications. You can also apply your own set of augmentation operations on data in the datastore by using the transform and combine functions. During training, the datastore randomly perturbs the training data for each epoch, so that each epoch uses a slightly different data set.

Image Processing Applications

Augment image data to simulate variations in the image acquisition. For example, the most common type of image augmentation operations are geometric transformations such as rotation and translation, which simulate variations in the camera orientation with respect to the scene. Color jitter simulates variations of lighting conditions and color in the scene. Artificial noise simulates distortions caused by the electrical fluctuations in the sensor and analog-to-digital conversion errors. Blur simulates an out-of-focus lens or movement of the camera with respect to the scene.

Common image preprocessing operations include noise removal, edge-preserving smoothing, color space conversion, contrast enhancement, and morphology.

If you have Image Processing Toolbox™, then you can process data using these operations as well as any other functionality in the toolbox. For an example that shows how to create and apply these transformations, see Augment Images for Deep Learning Workflows.

Processing TypeDescriptionSample FunctionsSample Output
Resize imagesResize images by a fixed scaling factor or to a target size

The original image is on the left . The resized image is on the right.

Warp imagesApply random reflection, rotation, scale, shear, and translation to images

From left to right, the figure shows the original image, the reflected image, the rotated image, and the scaled image.

Crop imagesCrop an image to a target size from the center or a random position

The image cropped from the center is on the left. The image cropped from a random position is on the right.

Jitter colorRandomly adjust image hue, saturation, brightness, or contrast

From left to right, the figure shows the original image with random adjustments to the image hue, saturation, brightness, and contrast.

Simulate noiseAdd random Gaussian, Poisson, salt and pepper, or multiplicative noise

The image with randomly added salt and pepper noise is on the left. The image with randomly added Gaussian noise is on the right.

Simulate blurAdd Gaussian or directional motion blur

The image with a Gaussian blur is on the left. The image with a directional motion blur is on the right.

Object Detection

Object detection data consists of an image and bounding boxes that describe the location and characteristics of objects in the image.

If you have Computer Vision Toolbox™, then you can use the Image Labeler (Computer Vision Toolbox) and the Video Labeler (Computer Vision Toolbox) apps to interactively label ROIs and export the label data for training a neural network. If you have Automated Driving Toolbox™, then you also use the Ground Truth Labeler (Automated Driving Toolbox) app to create labeled ground truth training data.

When you transform an image, you must perform an identical transformation to the corresponding bounding boxes. If you have Computer Vision Toolbox, then you can process bounding box data using the operations in the table. For an example that shows how to create and apply these transformations, see Augment Bounding Boxes for Object Detection. For more information, see Getting Started with Object Detection Using Deep Learning (Computer Vision Toolbox).

Processing TypeDescriptionSample FunctionsSample Output
Resize bounding boxesResize bounding boxes by a fixed scaling factor or to a target size

Crop bounding boxesCrop a bounding box to a target size from the center or a random position

Warp bounding boxesApply reflection, rotation, scale, shear, and translation to bounding boxes

Semantic Segmentation

Semantic segmentation data consists of images and corresponding pixel labels represented as categorical arrays.

If you have Computer Vision Toolbox, then you can use the Image Labeler (Computer Vision Toolbox) and the Video Labeler (Computer Vision Toolbox) apps to interactively label pixels and export the label data for training a neural network. If you have Automated Driving Toolbox, then you also use the Ground Truth Labeler (Automated Driving Toolbox) app to create labeled ground truth training data.

When you transform an image, you must perform an identical transformation to the corresponding pixel labeled image. If you have Image Processing Toolbox, then you can preprocess pixel label images using the functions in the table and any other toolbox function that supports categorical input. For an example that shows how to create and apply these transformations, see Augment Pixel Labels for Semantic Segmentation. For more information, see Getting Started with Semantic Segmentation Using Deep Learning (Computer Vision Toolbox).

Processing TypeDescriptionSample FunctionsSample Output
Resize pixel labelsResize pixel label images by a fixed scaling factor or to a target size

Crop pixel labelsCrop a pixel label image to a target size from the center or a random position

Warp pixel labelsApply random reflection, rotation, scale, shear, and translation to pixel label images

Lidar Processing Applications

Lidar Toolbox™ enables you to design, analyze, and test lidar systems. You can perform object detection and tracking, semantic segmentation, shape fitting, and registration. Raw point cloud data from lidar sensors requires basic processing before you can use them for these advanced workflows.

Lidar Toolbox provides tools to perform preprocessing such as downsampling, filtering, aligning, and extracting features from point cloud data. You can also augment and transform point clouds to increase the diversity of your training data.

Use Lidar Viewer (Lidar Toolbox) app to visualize, analyze and measure point cloud data. You can preprocess data by using the built-in preprocessing algorithms or import a custom algorithm. For more information, see Create Custom Preprocessing Workflow with Lidar Viewer (Lidar Toolbox).

You can create labeled ground truth training data by using the Lidar Labeler (Lidar Toolbox) app. For more information on automated labelling, see Automate Ground Truth Labeling for Vehicle Detection Using PointPillars (Lidar Toolbox).

Processing TypeDescriptionSample FunctionsSample Output
Clean and filter point cloud data

  • Downsample point cloud data using downsampling algorithms

  • Apply median filtering

  • Remove noise

Organize point cloudConvert point cloud into organized format, where you arrange the data as rows and columns according to the spatial relationship between the points 
size(ptCloudUnorg.Location)
ans = 1×2

       37879           3
ptCloudOrg = pcorganize(ptCloudUnorg,params);
size(ptCloudOrg.Location)
ans = 1×3

          64        1024           3
Create blocked point cloudsWhen your data is too large to fit into the memory, divide and process the point cloud as discrete blocks

Augment point cloud data

  • Apply a geometric transformation, such as random rotation, translation, shearing, and scaling

  • Randomly add bounding boxes to the training data

Signal Processing Applications

Signal Processing Toolbox™ enables you to denoise, smooth, detrend, and resample signals. You can augment training data with noise, multipath fading, and synthetic signals such as pulses and chirps. You can also create labeled sets of signals by using the Signal Labeler (Signal Processing Toolbox) app and the labeledSignalSet (Signal Processing Toolbox) object. For an example that shows how to create and apply these transformations, see Waveform Segmentation Using Deep Learning.

Wavelet Toolbox™ and Signal Processing Toolbox enable you to generate 2-D time-frequency representations of time series data that you can use as image inputs for signal classification applications. For an example, see Classify Time Series Using Wavelet Analysis and Deep Learning. Similarly, you can extract sequences from signal data to use as input for LSTM networks. For an example, see Classify ECG Signals Using Long Short-Term Memory Networks (Signal Processing Toolbox).

Communications Toolbox™ expands on signal processing functionality to enable you to perform error correction, interleaving, modulation, filtering, synchronization, and equalization of communication systems. For an example that shows how to create and apply these transformations, see Modulation Classification with Deep Learning.

You can process signal data using the functions in the table as well as any other functionality in each toolbox.

Processing TypeDescriptionSample FunctionsSample Output
Clean signals

  • Apply median filtering or moving average to signal

  • Remove polynomial trend

  • Resample signal to new fixed rate

Filter signals

  • Perform lowpass, highpass, and bandstop filtering of IIR and FIR signals

  • Design IIR and FIR filters

  • Apply IIR and FIR filters

Augment signals

  • Add white Gaussian noise to signal using Communications Toolbox

  • Adjust time information of the signal, and perform multipath fading using Communications Toolbox

  • Add synthetic chirps and waveforms

  • awgn (Communications Toolbox)

  • chirp (Signal Processing Toolbox), square (Signal Processing Toolbox), rectpuls (Signal Processing Toolbox), sawtooth (Signal Processing Toolbox)

Create time-frequency representationsCreate spectrograms, scalograms, and other 2-D representations of 1-D signals

  • pspectrum (Signal Processing Toolbox), xspectrogram (Signal Processing Toolbox)

  • fsst (Signal Processing Toolbox), ifsst (Signal Processing Toolbox)

  • stft (Signal Processing Toolbox), istft (Signal Processing Toolbox)

  • cwt (Wavelet Toolbox)

Extract features from signalsEstimate instantaneous frequency and spectral entropy

Audio Processing Applications

Audio Toolbox™ provides tools for audio processing, speech analysis, and acoustic measurement. Use these tools to extract auditory features and transform audio signals. Augment audio data with randomized or deterministic time scaling, time stretching, and pitch shifting. You can also create labeled ground truth training data by using the Signal Labeler (Signal Processing Toolbox) app. You can process audio data using the functions in this table as well as any other functionality in the toolbox. For an example that shows how to create and apply these transformations, see Augment Audio Dataset (Audio Toolbox).

Audio Toolbox also provides MATLAB® and Simulink® support for pretrained audio deep learning networks. Locate and classify sounds with YAMNet and estimate pitch with CREPE. Extract VGGish or OpenL3 feature embeddings to input to machine learning and deep learning systems. The Audio Toolbox pretrained networks are available in Deep Network Designer. For a YAMNet example, see Adapt Pretrained Audio Network for New Data Using Deep Network Designer.

Processing TypeDescriptionSample FunctionsSample Output
Augment audio dataPerform random or deterministic pitch shifting, time-scale modification, time shifting, noise addition, and volume control

Extract audio featuresExtract spectral parameters from audio segments

Processed output:

ans = struct with fields:
                mfcc: [1 2 3 4 5 6 7 8 9 10 11 12 13]
           mfccDelta: [14 15 16 17 18 19 20 21 22 23 24 25 26]
      mfccDeltaDelta: [27 28 29 30 31 32 33 34 35 36 37 38 39]
    spectralCentroid: 40
               pitch: 41

Create time-frequency representations

  • Create mel spectrograms and other 2-D representations of audio signals

  • Prepare audio signals to feed to pretrained deep learning networks

Text Analytics

Text Analytics Toolbox™ includes tools for processing raw text from sources such as equipment logs, news feeds, surveys, operator reports, and social media. Use these tools to extract text from popular file formats, preprocess raw text, extract individual words or multiword phrases (n-grams), convert text into numerical representations, and build statistical models. You can process text data using the functions in this table as well as any other functionality in the toolbox. For an example showing how to get started, see Prepare Text Data for Analysis (Text Analytics Toolbox).

Processing TypeDescriptionSample FunctionsSample Output
Tokenize text Parse text into words and punctuation

Original:

"A few tree limbs greater than 6 inches down on HWY 18 in Roseland."

Processed output:

15 tokens: A few tree limbs greater than 6 inches down on HWY 18 in Roseland .

Clean text

  • Remove variations in word forms and case

  • Remove punctuation

  • Remove stop words, short words, and long words

Processed output:

15 tokens: a few tree limb great than 6 inch down on hwy 18 in roseland .

14 tokens: a few tree limb great than 6 inch down on hwy 18 in roseland

8 tokens: few tree limb great inch down hwy roseland

See Also

| | | | |

Related Topics