Digital Signal Processing Reference
In-Depth Information
Image Components Extraction (IE) layer, (ii) Visual Contents Extraction (CE) layer,
(iii) Visual Content Matching (CM) layer and (iv) Objects Recognition (OR) layer.
They are hierarchically organized as the illustration shown in Figure 2, capturing an
image as input and the identified objects as output.
Fig. 2.
Hierarchical Objects Recognition Model (Left Column) mapping to the High-level
human vision object recognition stages according to cognitive neuroscience of visual object
recognition process (Right Column)
In the IE layer, multiple regions from different parts of an input image are required
to be identified. The regions are different in size and located randomly around the
image. We introduced a Random Location Subwindows (RLS) mechanism for this
operation. The mathematic model is explained in details next section. The next layer
that linked directly to IE layer is CE layer. The operation in CE layer is to extract the
low-level visual information from these multiple targeted regions. The information is
encoded into different type of VD, respective to what have been used in OF conceptu-
alization. In the CM layer, the encoded visual information will then be compared to
the pre-registered information in OF KB. The comparison of the extracted VD and all
the VD associated with the concepts in OFKB will drive every concept marked with a
ranking value. The concepts that record higher similarity values in their VD compari-
son will score a higher ranking value. Thus, in OR layer, all the marked concepts will
be analysed. A decision will be derived from the analysis to determine a detected
concept.
This multi-layer operations in-combination enable objects in an image being identi-
fied, simulates the object recognition states in human vision, as shown in Figure 2.
Various image components being extracted from multiple parts of an image without
the needs of prior object segmentation process. Together with the low-level image
processing that is performed on these specific parts. They are analogue to the process
of a human is getting visual attention on certain parts of the image. In addition, the
VD comparison in OF KB process as the similar analogy to the neuron firing mecha-
nism for information searching in human memory. If the comparison of the features
