Digital Signal Processing Reference
In-Depth Information
Cognitive Semantic Model for
Visual Object Recognition in Image
Sieow Yeek Tan and Dickson Lukose
Artificial Intelligence Center, MIMOS Berhad,
Technology Park Malaysia, Bukit Jalil, Kuala Lumpur, Malaysia
{tan.sy,dickson.lukose}@mimos.my
Abstract.
The paper presents a hierarchical semantic model to perform object
recognition in 2D images using cognitive neuroscience vision process. The pro-
posed model contains two vital parts: Object-Features (OF) Conceptualization
and Concept Recognition (CR). The model facilitates combination of multiple
visual descriptors in the OF conceptualization and the CR process. The model
comprises four major operation layers: Image Components Extraction (IE
Layer), Visual Content Extraction (CE Layer), Visual Content Matching (CM
Layer) and Object Recognition (OR Layer), arranged hierarchically from bot-
tom (IE Layer) to top (OR Layer). The OR layer incorporates Multi-Level
Thresholds technique, which defines various threshold values to control and fi-
nalize CR process. The experiments performed involved two types of visual de-
scriptors: Color and Edge Directivity Descriptor (CEDD) and Fuzzy Color and
Texture Histogram (FCTH). They were carried out using 9 set of images data-
set. Different threshold values were tested to validate the feasibility and accu-
racy of the proposed model. Intensive empirical assessment has been performed
and the results are promising.
Keywords:
Object Recognition, Image Understanding, Multiple Feature
Descriptors, Hierarchical Semantic Model.
1
Introduction
Recognizing the content of image in term of objects has been a challenging problem
in computer vision since the past decades. Some of the popular methods in addressing
this problem are based on hierarchical approaches, which is inspired by the hierarchi-
cal nature of visual cortex [1-2]. The performance of hierarchical approaches has been
always outperforming the single-template object recognition systems [3-4]. Neverthe-
less, in detecting certain type of objects, for instance faces, cars, pedestrians; single-
template systems are still able to exhibit excellent performance [5-7].
We believe biological inspired hierarchical architecture with supervised learning
will be a powerful method in solving visual object recognition problem. It can surpass
some of the general purpose machine learning methods. Studies [1-2] have shown a
concrete positive results in adopting such approach. In this paper, we propose a novel
model, which is referring to human vision object recognition stages, according to
