Computer Vision News - August 2016

It’s more than a deep learning package, it’s an architecture! Last November, Google released the open source TensorFlow™ software library. TensorFlow has become synonymous with "deep learning" in the community. To quote Wikipedia: “ TensorFlow is an open source software library for machine learning in various kinds of perceptual and language understanding tasks ”. However, TensorFlow is not only about deep learning or machine learning. TensorFlow can also be used as a library for numerical computation using data flow graphs. TensorFlow allows you to create a graph of variables and operations. The variables are the edges of the graph (also called Tensors), representing data, which could be a number, a string, a matrix, and so on. The operations are the notes of the graph. These represent mathematical operations, functions, transformations. One example of the benefits of using the TensorFlow architecture is that it allows you to deploy computation in a parallel fashion to one or more CPUs, GPUs, a mobile device or in a cluster, all this with a single API. Yes, it is much more than a deep learning package . In this month's tool section we will demonstrate how one can use the TensorFlow architecture to do some computer vision tasks. For this illustration, we use the Hough Line Transform . The code is written in Python and uses the OpenCV library. In a nutshell, the Hough Transform is a feature extraction technique used to find instances of objects of a certain class of shapes (e.g. lines, circles). Specifically, the Hough Line Transform keeps track of the intersection between curves of every point in the image. When the number of intersections is above some predefined threshold, it declares it as a line with the parameters . 8 Computer Vision News Tool TensorFlow - Part 1 Tool In our demonstration, we will search for the line in the Sudoku image at the right: Let’s see the code for this. First some initialization and loading of relevant libraries ( OpenCV , Numpy , MatplotLib ) :