UNI-MB - logo
UMNIK - logo
 
Nacionalni inštitut za biologijo in BF, Oddelek za biologijo, Ljubljana - vsi oddelki (BFBNIB)
  • Polyp counting made easy: two stage scyphistoma detection for a computer-assisted census in underwater imagery
    Vodopivec, Martin, 1979- ...
    Moon jellyfish have a complex life cycle that alternates between a pelagic free-swimming medusa and an attached polyp. Understanding the population dynamics of the latter is of vital importance for ... understanding the abundance of the bloom forming medusa stage. In order to efficiently analyze a large corpus of underwater photographs taken during a 3-year survey of Aurelia polyps in situ, we have developed software that is capable of automatic detection and counting of polyps present in the image. We combined state-of-the-art computer vision and machine learning methods in a two-stage approach: first, potential scyphistoma candidates are identified by an Aggregate Channel Features (ACF) detector, and then verified by a combination of Support Vector Machine (SVM) classifier and features, extracted using a Convolutional Neural Network (CNN). This approach was tested on several photographs, each containing roughly a thousand polyps, and its accuracy was compared to the accuracy and variance of a manual census performed by multiple experts using two different manual counting methods. The computer-assisted approach was shown to drastically reduce the effort and time spent counting polyps in an image at a minimal reduction in accuracy, thus enabling processing of much larger datasets. The developed program can be used for the detection of different scyphistoma on virtually any substrate; and being based on machine learning, it is also highly flexible and can be further improved through user interaction. The program itself will be made publicly available.
    Vrsta gradiva - prispevek na konferenci
    Leto - 2016
    Jezik - angleški
    COBISS.SI-ID - 3915343