Page 59 - Flytxt
P. 59
contains some specific object, feature or corresponding to each image. DAISY
an activity of interest. features associated with key points
Excitingly, Flytxt is now pitching in to within images as the local feature
solve this problem. The Data Science descriptor (similar to SIFT features) and
R&D team at Flytxt has released an histogram of oriented gradients (HOG)
end-to-end scene recognition pipeline corresponding to an entire image is used
consisting of feature extraction, as a global descriptor.
encoding, pooling and classification. As images vary in view point, scale,
The primary objective of this project is to orientation, illumination and occlusion
clearly outline the practical level of objects, extracting robust
implementation of a basic scene- features (such as DAISY, SIFT, HOG etc.)
recognition pipeline having reasonable to represent them is critical for building
accuracy, using conventional computer an effective image classification model.
vision techniques (without applying deep As the number of key points vary across
learning techniques) in Python, using images, multiple DAISY descriptors would
open-source libraries. exist for each image. We use a bag-of-
How the Flytxt model works visual-words concept to encode each
To enable accurate recognition, Flytxt image as a histogram of dimensionality
‘K’ (where K is the vocabulary size or the
R&D team simultaneously utilizes
global feature descriptors as well as number of possible “visual words”).
local feature descriptors from images Clustering is then used to group DAISY
to form a hybrid feature descriptor features to form the “visual words”
Fig - 3-3-1 Source: Flytxt
INSIGHTZ - VOLUME 03, 2018 59
