Synergistic Fusion of Shape and Colour for RPDetRec: Automating Retail
Product Detection and Recognition
Abstract
Detection and Recognition of on-shelf and off-shelf retail products is
challenging in computer vision because of an extremely high number of
different products. Some of the existing approaches extracted text
information to recognize the retail products. However, these approaches
work effectively only if the text information, such as product name,
price, and expiration date, is explicitly visible in the image. In this
paper, a novel image processing framework is implemented to detect and
recognize retail products using shape and colour features. This
framework comprises of three modules: illumination-invariant
representation, colour-based object identification, and shape-based
object identification. The images acquired from retail stores may
contain reflective effects such as shadows and illuminations because
they are captured in an uncontrolled environment that can be eliminated
using a three-step process defined in the first module. In the second
module, the salient region detection and mean shift segmentation
algorithms are utilized to capture the location of the product region
based on colour features from a rectified image. After removing
backgrounds, the segmented region of interest containing retail products
is given as input to the final module. In the final module, a Contour
Line Inflection Arc Extraction (CLIAE) algorithm is proposed to detect
and recognize retail products using shape features. Finally, the Speeded
Up Robust Features (SURF) algorithm is employed to perform the object
matching between the known object and the target object. The performance
and robustness of the proposed framework are evaluated using two
benchmark datasets, GroZi-3.2k and Grocery Products (GP). However, the
proposed framework outperforms both combined features-based object
recognition algorithms and state-of-the-art retail product recognition
algorithms by achieving a recognition accuracy of 95.5% and 95.8% on
GroZi-3.2k and GP datasets, respectively. The usefulness of each module
in the framework is also demonstrated by conducting a variety of
experiments on benchmark datasets.