A Comparison of Canny Edge Detection Implementations with Hadoop

Edge detection plays a large role in digital image processing; it allows for easier identification of objects within the image from a human perspective and also opens the door for automated object detection via machine learning. One of the common edge detection algorithms, developed by John F. Canny in 1986, is a multi-stage algorithm making use of gradients within the image to calculate potential edges. Hadoop is a Java library built on the MapReduce framework and was designed to be used for distributed processing of big data projects. We, instead, choose to adapt the Canny edge detection algorithm to run via Hadoop using two methods: a streaming Python implementation and a Java implementation to compare their run times and to determine whether or not using Hadoop for such a problem is desirable over the classic sequential implementation. In Sect. 1, we introduce edge detection, explaining what it is and why it is important. In Sect. 2, we explore Canny’s algorithm. In Sect. 3, we then explain our methodology of parallelizing the sequential code and the issues which arise during the process. In Sect. 4, we display the results of our implementations. In Sect. 5 we conclude with a possible explanation of why our code does not perform as well as we had anticipated; due to memory limitations of our specific Hadoop cluster, the implementations created do not perform well for processing typical images, but do allow for the processing of very large images. We also explore possibilities for future work.