Dynamic Color Distribution Entropy (D-CDEN)

• The paper introduces D-CDEN, a method that augments traditional color histograms by encoding both color frequency and spatial entropy of connected pixel neighborhoods.
• It employs an 8-connected neighborhood detection and union-find merging technique to calculate entropy, eliminating arbitrary ring-based parameters.
• Experimental results on datasets like SIMPLIcity and Caltech-101 demonstrate a 5–10% precision improvement over static methods in image retrieval tasks.

Dynamic Color Distribution Entropy of Neighborhoods (D-CDEN) is a color-feature extraction method designed to augment traditional color histograms by encoding not only the frequency but also the spatial arrangement and dispersion of colors within an image. Unlike conventional global histograms or static ring-based descriptors, D-CDEN incorporates data-driven spatial analysis by detecting and quantifying the “neighborhoods” of connected pixels for each color bin, producing a joint feature that describes both color prevalence and its spatial entropy (Alamdar et al., 2012).

1. Principles and Objectives

D-CDEN was introduced to address the primary deficiency of color histograms: their insensitivity to spatial distribution. In contrast to the original Color Distribution Entropy (CDE), which measures spatial dispersion via a set of arbitrary concentric rings about bin centroids, D-CDEN constructs neighborhoods as naturally connected components within each bin. This yields the following objectives:

• Encode the compactness versus dispersion of every color bin by analyzing data-driven neighborhoods.
• Remove dependence on arbitrary, global parameters such as number or width of rings.
• Enhance the discriminative power of image retrieval, clustering, and indexing systems by providing a richer color spatial signature.

Neighborhood detection responds adaptively to actual image content and topology, thereby overcoming several representational limitations inherent in previous approaches.

2. Mathematical Formulation

Let an image be quantized into $n$ color bins in a uniform HSV space partition (e.g., $n=32$ for $8 \times 2 \times 2$ quantization). For each bin $i$, define:

• $A_i$: Set of all pixels assigned to bin $i$
• $|A_i|$: Total pixel count in bin $i$

Neighborhood extraction proceeds by scanning the image (typically $128 \times 128$ pixels) top-to-bottom and left-to-right, grouping pixels of equal bin label into 8-connected components. The resulting data-driven set of neighborhoods for bin $i$ is $n=32$0, where $n=32$1 is the number of distinct neighborhoods in bin $n=32$2.

The Normalized Spatial Distribution Histogram (NSDH) for bin $n=32$3 is:

$n=32$4

Spatial dispersion is quantified using Shannon entropy:

$n=32$5

High $n=32$6 indicates a color that is scattered across many small neighborhoods; low $n=32$7 reflects color concentrated into large contiguous regions.

The resulting $n=32$8-dimensional D-CDEN feature vector:

$n=32$9

where $8 \times 2 \times 2$0 (optionally normalized by total pixel count).

3. Algorithmic Pipeline and Implementation Specifics

D-CDEN computation entails the following stages (see (Alamdar et al., 2012) for full pseudocode):

1. Resize input RGB image to $8 \times 2 \times 2$1 to suppress noise and limit minute regions.
2. Transform to HSV color space and uniformly quantize into $8 \times 2 \times 2$2 bins.
3. Initialize per-pixel bin labels and per-bin neighborhood lists.
4. For each pixel (scanned in raster order), examine up, up-right, left, and up-left neighbors of same bin label among previously visited pixels:
   • If no neighbor matches, initialize a new neighborhood ID.
   • If exactly one match, assign pixel to that neighborhood.
   • If several matches, merge all neighboring IDs into one and assign.
5. For each bin $8 \times 2 \times 2$3, calculate $8 \times 2 \times 2$4 and $8 \times 2 \times 2$5 as described above.
6. Output the $8 \times 2 \times 2$6 feature vector.

Key implementation details include 8-connectivity (only to already scanned pixels), online merging of neighborhoods to accommodate growing and merging regions, and optional normalization of histogram counts.

4. Computational Complexity

Considering $8 \times 2 \times 2$7 as the number of pixels and $8 \times 2 \times 2$8 as the number of color bins:

Method	Complexity	Notes
Plain histogram	$8 \times 2 \times 2$9	Per-pixel bin assignment only
CDE (static rings)	$i$0	$i$1 = number of rings; global parameter introduces cost
D-CDEN	$i$2	Neighborhood assignment/merging via union-find; entropy calculation bounded by $i$3

The D-CDEN approach thus maintains near-linear complexity with manageable overhead relative to histogram-only or static-parameter methods, while avoiding arbitrary ring choices entirely.

5. Experimental Evaluation

D-CDEN was empirically tested on:

• SIMPLIcity dataset: 1,000 images over 10 categories (100 images each)
• Caltech-101: 6,384 images across 70 categories

Queries were performed using 100 images from selected categories in SIMPLIcity (Buildings, Buses, Flowers) and 50 randomly chosen images per run in Caltech-101. The similarity or distance between images was computed with a combined Vector-Space Model (VSM):

$i$4

where each cosine term denotes dot-product similarity for histogram, entropy, and normalized neighborhood-count components.

D-CDEN achieved superior retrieval performance compared to Improved Color Distribution Entropy (I-CDE) at every recall level tested. For example, on the "Buildings" subset and Recall = 0.5, D-CDEN precision was approximately 0.62 versus I-CDE's 0.53. Gains of 5–10 percentage points were consistently observed on Buses, Flowers, and aggregate scores for Caltech-101 (Alamdar et al., 2012).

6. Qualitative and Theoretical Insights

D-CDEN improves semantic clustering of retrieval results. For instance, a bus-category query yields top results that consistently depict buses with congruent viewpoints, whereas I-CDE returns several irrelevant or mismatched images. The method's sensitivity to actual pixel connectivity allows it to distinguish images where a color bin appears in multiple discrete blobs from those where it is concentrated, penalizing the former as more spatially fragmented. This enables more reliable discrimination of visually coherent versus scattered patterns.

A plausible implication is that neighborhood-level entropy, as a first-order spatial statistic, provides more meaningful information about color-region structure than purely global or ring-based metrics.

7. Limitations and Prospective Extensions

D-CDEN neighborhood detection may be affected by quantization artifacts, resulting in small, potentially spurious neighborhoods. Additionally, the current method encodes only first-order statistics—the entropy of neighborhood sizes—while neglecting higher-order relationships such as spatial distribution of neighborhoods relative to each other. The proposed future directions include:

• Integration with texture or shape descriptors for richer multi-modal feature sets
• Incorporation of weighting schemes akin to I-CDE for enhanced perceptual matching
• Exploration of higher-order spatial statistics to better capture complex region arrangements (Alamdar et al., 2012)

These enhancements could further improve robustness to noise and discriminative capacity for challenging retrieval and recognition problems.