Download presentation
Presentation is loading. Please wait.
1
Data Preparation for Deep Learning
2
Why do we need big data? Supervised machine learning is practical & reliable Personal devices are broadcasting a colossal amount of information every second Deep learning is becoming an extremely popular technique for machine learning tasks
4
Image Processing Computationally intensive
“The sum is greater than the parts” A single image is a huge set of data 3 RGB channels, with dimensions 1024x768: 2,359,296 bytes per image 1,800,000 digital images uploaded everyday: 4,246,732,000,000 bytes per day
5
Datasets for Deep Convolutional Neural Networks
Requirements: Datasets should contain a minimum of 4,000 images per image class All image classes should be realistically distributed All images should be cropped to their region of interest Duplicates must be removed (to avoid overfitting) All images must be downscaled to a size of n*n All images must be normalized
6
What is a feature vector?
A vector that describes a set of “features” from a piece of data. One tuple describes one entity (in our case, one image), and each row within that tuple pertains to a measurement that describes our entity. In this case, we have color, sepal width, sepal height, petal width, and petal height. For each feature that we use to describe our entity, our feature vector increases in “dimensionality”. purple 2.96cm 3.14cm 1.98cm 2.11cm Image to 5-entry tuple “Feature vector has dimensionality of 5”
7
Clustering Similar Images for Class Distribution
Clustering (or Lloyd’s algorithm) can be thought of as comparing feature vectors, and then organizing these feature vectors according to their similarity. If two feature vectors have very similar values, they will be placed into the same cluster. If we cluster images with similar features, we can ensure an accurate image class distribution by observing the amount of images belonging to each cluster.
8
Clustering Similar Images for Class Distribution
9
Drawbacks of Lloyd’s Algorithm
For each feature within our feature vector, we increase the dimensionality of our problem by 1. In practical applications, we use d-dimensional feature vectors, where d > 3. As mentioned in MMDS, clustering suffers from the “curse” of dimensionality, where most points are far apart from one another in high-dimensional space: We can overcome this by using dimensionality reduction techniques, such as principal component analysis.
10
Overview of Principal Component Analysis and Dimensionality Reduction
If we revisit our iris example, we can make a simple observation: Sepal width and sepal height appear to be closely correlated. Petal width and petal height appear to be closely correlated. If this holds true for a large number of samples, then we can reduce the dimensionality of our feature vector without losing much information: purple 2.96cm 3.14cm 1.98cm 2.11cm Image to 5-entry tuple “Feature vector has dimensionality of 5”
11
Overview of Principal Component Analysis and Dimensionality Reduction
If we revisit our iris example, we can make a simple observation: Sepal width and sepal height appear to be closely correlated. Petal width and petal height appear to be closely correlated. If this holds true for a large number of samples, then we can reduce the dimensionality of our feature vector without losing much information: purple 2.96cm 1.98cm Image to 3-entry tuple “Feature vector has dimensionality of 3”
12
Efficiency of Lloyd’s Algorithm
Since Lloyd’s algorithm is a heuristic (it doesn’t provide an exact solution), it’s hard to determine an exact function for its time complexity. Most analyses describe it as O(n * k * d * i). In practice, this family of algorithms is extremely quick and accurate, as long as we don’t use highly- dimensional feature vectors.
13
What is an image descriptor?
An image descriptor is a numerical measure of an image. Colour Histogram of an iris in LAB colour-space:
14
The problem with our clustering approach
Clustering a feature vector is trivial, but calculating that feature vector is not. Calculating visual feature descriptors on-the-fly are computationally expensive. We need to find a way to minimize the amount of image data that is being processed.
15
The solution: Find regions of interest in our image
If we can crop out the portion of the image that we’re interested in, we can massively reduce the number of pixels that we need to process. This will also give us more meaningful results when we cluster our images.
16
Saliency in Statistical Scene Analysis
Saliency can be informally defined as the portion of an image that “stands out”. Natural images follow something known as the 1/f law. This means that in a natural, “ambient” scene, the amplitude of the averaged fourier spectrum follows a common distribution.
17
Saliency in Statistical Scene Analysis
By subtracting the natural distribution from the distribution of our image, we’re left with a metric that indicates where unnatural “salient” regions of our image occur. We call this measure the “spectral residual” of an image, and this spectral residual is then used to create a “saliency map”.
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.