Modern image collections cannot be processed e ciently on one computer due to large collection sizes and high computational costs of modern image processing algorithms. Hence, image processing often requires distributed computing. However, distributed computing is a complicated subject that demands deep technical knowledge and often cannot be used by researches who develop image processing algorithms. The framework is needed that allows the researches to concentrate on image processing tasks and hides from them the complicated details of distributed computing. In addition, the framework should provide the researches with the familiar image processing tools. The paper describes the extension to the MapReduce Image Processing (MIPr) framework that provides the ability to use OpenCV in Hadoop cluster for distributed image processing. The modi ed MIPr framework allows the development of image processing programs in Java using the OpenCV Java binding. The performance testing of created system on the cloud cluster demonstrated near-linear scalability.
Image processing is one of the important tasks in many areas of research such as medical imaging, remote sensing, astronomy, Internet analysis, etc. The industry also widely uses image processing.
Nowadays, image processing often requires distributed computing. The sizes
of modern image collections, for example, the Yahoo 100 million creative
commons ickr images for research [
However, distributed computing is a complicated subject that requires deep technical knowledge. Many scientists, who developed image processing algorithms, does not have the quali cation required to use distributed computing e ciently. They need an image processing framework that hides the complicated details of distributed computing from the application developers and allows them to concentrate on image processing tasks and algorithms.
Previously, we have developed the MapReduce Image Processing framework
(MIPr) [
However, providing the application developers with a convenient framework is not enough for e cient distributed image processing. The developers must be able to use image processing tools that they already know together with the framework. In such case, the developers can quickly create software for distributed computing framework, or even use their existing software. Otherwise, the developers will have to implement image processing algorithms from scratch, which require a lot of time and labor.
In this paper we describe an approach for using the popular open source
computer vision library OpenCV [
Several systems can be used for image processing in Hadoop. HIPI (Hadoop
Image Processing Interface) [
The image processing cloud with the OpenCV support was built at the Prairie
View A&M University (PVAMU)[
Libraries Java API
C++ API
Python API Map-Reduce Drivers OpenCV Map-Reduce
Driver
Image processing
API
I/O OpenCV I/O tools: - Input Format - Combine File Input Format - Record Reader - Output Format - Record Writer
Serialization
Mat writable
Core components 3 Integration of OpenCV Library into the MIPr
Framework The MIPr framework has an open architecture, which allows adding new image representation based on various libraries. The extended architecture of MIPr with the support of OpenCV library is presented in Fig 1. The extensions include an OpenCV-based image representation in the Hadoop internal format, input/output tools for inclusion of OpenCV-based images into the Hadoop data work ow, and MapReduce Driver to execute Hadoop programs with the OpenCV support.
OpenCV uses the Mat class as an image container, and most of OpenCV image processing functions accept Mat as an argument. Hence, the Mat class was chosen to build the OpenCV-based image format in MIPr. Similar to other MIPr image representation formats, the Hadoop Writable interface was used for serialization. By combining these two components, the MatWritable image format was created, which is suitable for use as a value in MapReduce programs.
For reading images from the Hadoop Distributed File System (HDFS) into the memory for processing, two types of InputFormat for OpenCV have been developed: OpenCVFileInputFormat and OpenCVCombineFileInputFormat. OpenCVFileInputFormat reads one image at a time and is suitable for big images (larger than the HDFS block size). In contrast with it, OpenCVCombineFileInputFormat reads several images from the same HDFS block in one operation and is suitable for processing large amount of small les. Both InputFormats produce pairs < le name, MatWritable> (OpenCVCombineFileInputFormat can produce several pairs if the HDSF block contains several images). To actually read an image from HDFS and present it in the MatWritable format, the OpenCVRecordReader was developed, which is used both by the OpenCVFileInputFormat and OpenCVCombineFileInputFormat.
For writing the OpenCV-based images to HDFS after processing, special implementations of OutputFormat and RecordWriter, which are capable of working with MatWritable, were developed. In contrast to the traditional Hadoop approach, the created components preserve the original names of image les. 3.3
MapReduce driver is intended to execute the MapReduce program on the cluster. The developed OpenCV MapReduce driver sets up the appropriate InputFormat and OutputFormat, de nes the MapReduce values type as a MatWritable, loads OpenCV native C++ library, and executes the job on the Hadoop cluster. To distribute the binary OpenCV C++ library to the cluster nodes, the Hadoop distributed cache is used.
The MIPr extensions use the OpenCV Java binding. In contrast to the PVAMU cloud, MIPr utilizes the standard Hadoop Mappers and Reducers for image processing. Hence, only Java is supported for developing image processing algorithms in MIPr using OpenCV. 4
In order to estimate the performance and scalability of the proposed approach, two computational experiments with the modi ed MIPr implementation have been conducted. The Microsoft Azure HDInsight cluster in the Microsoft Azure cloud platform have been used for the experiments. The cluster consists of 18 nodes: 2 head nodes and 16 computational nodes. The con guration of the nodes is presented in the Table 1.
The MIRFLICKR-1M image collection (one million of images downloaded
from Flickr, 118GB total size) [
The second experiment was aimed at estimating how the system scales with the increasing volume of images. For this purpose, the whole cluster was used to process varying number of images. The results are presented in Fig. 3 (logarithmic scale). For small datasets ( the number of images is less than 10000), there is no di erence in processing time due to the cluster overhead. However, the scalability for large image collection processing is near-linear.
Near-linear scalability of the modi ed MIPr implementation is an expected result, because images are processed independently. Processing time,
sec 10000 1000 100 10 1
Img2Gray
FaceDetection
CannyEdgeDetection 1000 10000 100000
1000000
Number of images in the collecton The approach to use OpenCV library for distributed image processing on a Hadoop cluster was presented. The approach is based on the previously developed MIPr framework for image processing in Hadoop. The MIPr framework was extended to include OpenCV-based image representation in the Hadoop internal format, tools for reading images from HDFS and writing processed images back to HDFS, and drivers for MapReduce jobs with OpenCV support. The extension allows to develop image processing programs in Java using OpenCV Java bindings.
The performance of the modi ed MIPr implementation was evaluated. For large clusters and image datasets the scalability is near-linear. Poor performance of single node cluster and small dataset processing is caused by Hadoop overhead.
MIPr framework is open sourced and available for free at https://github. com/sozykin/mipr.
Future work include porting MIPr to Apache Spark [
Acknowledgements. The reported study was supported by RFBR (research project No 14-07-31324 mol a) and by Microsoft Azure for Research Russia Initiative. Our work was performed using the \Uran" supercomputer from Institute of Mathematics and Mechanics UrB RAS.