An Efficient Content-Based Image Retrieval System for Clothing Images Using Color–Texture Fusion and K-means Clustering

An Efficient Content-Based Image Retrieval System for Clothing Images Using Color–Texture Fusion and K-means Clustering


1.    Introduction
In recent years, the rapid advancement of digital imaging technologies, high-speed internet, and large-scale storage systems has led to an unprecedented growth in the volume of digital images. These images are widely used across various domains such as healthcare, surveillance, multimedia systems, and especially e-commerce platforms [1, 2, 3]. In the context of online shopping, millions of clothing images are uploaded daily, making it increasingly difficult for users to efficiently locate items that match their preferences.
Traditionally, image retrieval systems have relied on text-based annotations, where images are manually labeled using descriptive keywords. For example, a clothing image might be labeled as “red dress” or “formal shirt.” However, this approach suffers from several significant limitations. First, manual annotation is highly labor-intensive and time-consuming, particularly for large datasets. Second, it introduces subjectivity, as different annotators may describe the same image differently. For instance, one user may label an image as “dark red,” while another may describe it as “maroon,” leading to inconsistency in retrieval results [4]. Furthermore, textual descriptions often fail to capture low-level visual details such as texture patterns or subtle color variations, which are crucial in clothing applications.
To overcome these limitations, Content-Based Image Retrieval (CBIR) systems have been developed. Unlike text-based methods, CBIR systems analyze the visual content of images directly, extracting features such as color, texture, and shape [5]. These features provide a more objective and consistent representation of image content. For example, the color distribution of a dress can be represented using histograms, while the texture of the fabric can be characterized using statistical measures.
Among the various visual features, color and texture play a particularly important role in clothing image retrieval. Color is often the first attribute perceived by users and is commonly used in search queries (e.g., “blue jeans” or “black jacket”). On the other hand, texture provides essential information about the material and pattern of the clothing, such as whether the fabric is smooth, rough, striped, or patterned. For instance, two garments may share the same color but differ significantly in texture (e.g., silk versus denim), which directly affects user preference and perception [7]. Therefore, relying on a single feature is often insufficient for accurate retrieval.
Despite the effectiveness of CBIR systems, several challenges remain. One of the primary challenges is the computational cost associated with searching large-scale image databases. In a naive implementation, a query image must be compared with every image in the database, resulting in high processing time and reduced system efficiency. This issue becomes more critical as the size of the dataset increases, particularly in real-world applications such as online fashion stores [8].
Another challenge is the semantic gap, which refers to the discrepancy between low-level visual features (e.g., color histograms) and high-level human perception (e.g., “elegant dress” or “casual wear”). While low-level features can describe images mathematically, they may not fully capture the semantic meaning perceived by users. Recent studies attempt to address this issue through deep learning and multimodal approaches that combine visual and textual information. However, such methods often require large labeled datasets and high computational resources.
To address these challenges, recent research has focused on hybrid approaches that combine multiple features and incorporate efficient indexing or clustering techniques. Feature fusion, which combines different types of visual features, has been shown to significantly improve retrieval accuracy by capturing complementary information [9]. At the same time, clustering techniques such as K-means can be used to partition the dataset into smaller groups, thereby reducing the search space and improving retrieval speed [16].
In this context, this research proposes an efficient CBIR system for clothing image retrieval that integrates color and texture features using a weighted fusion strategy. The system further employs K-means clustering to reduce computational cost by limiting the search to the most relevant cluster rather than the entire dataset. This approach aims to achieve a balance between accuracy and efficiency, making it suitable for practical applications such as online fashion recommendation systems.
The main contributions of this research can be summarized as follows:
•    Developing a CBIR system that combines color and texture features for improved representation 
•    Applying K-means clustering to reduce retrieval time 
•    Evaluating the system using standard performance metrics such as precision, recall, and retrieval time 
•    Demonstrating the effectiveness of the proposed method in clothing image retrieval scenarios 
Overall, this study contributes to the growing body of research on efficient CBIR systems by providing a practical and computationally efficient solution that bridges the gap between accuracy and scalability.

2.    Literature Review / Background
Content-Based Image Retrieval (CBIR) has been extensively studied over the past decades, evolving from simple feature-based approaches to more advanced hybrid and deep learning methods. This section reviews the most relevant studies, focusing on feature extraction, feature fusion, and search optimization techniques used in image retrieval systems, particularly for clothing images.
2.1.     Early CBIR Systems (Color-Based Approaches)
Early CBIR systems primarily relied on color features due to their simplicity and effectiveness in representing image content. One of the pioneering works in this field is the QBIC system proposed by Michael Flickner et al. (1995), which utilized color histograms and spatial indexing techniques to retrieve images based on visual similarity [5]. Similarly, Jeffrey Hafner et al. (1995) introduced an efficient color histogram indexing method using R-tree structures to reduce search space [12].
Although these approaches achieved promising results, they were limited by their reliance on color alone. For example, images with similar color distributions but different textures or shapes could not be effectively distinguished.
2.2.      Multi-Feature-Based Approaches
To overcome the limitations of single-feature systems, researchers began integrating multiple visual features such as color, texture, and shape. Eric Hsu et al. (2011) proposed a clothing image retrieval system that combines color, texture (using Gabor filters), and shape (using SIFT features), demonstrating significant improvement in retrieval accuracy [15].
Similarly, Manish Gupta et al. (2018) combined color and texture features to enhance retrieval performance. Their results showed that integrating complementary features can significantly improve precision, especially in clothing datasets where visual appearance is complex [20].
Recent studies further confirm the importance of feature fusion. For example, Wang et al. (2024) demonstrated that decoupling and combining color and texture features leads to better discrimination in fabric image retrieval tasks [27]. These findings highlight the effectiveness of multi-feature approaches in addressing the limitations of individual features.
2.3.     Search Space Reduction and Clustering Techniques
As image databases grew larger, computational efficiency became a critical issue. To address this challenge, researchers introduced indexing and clustering techniques to reduce the search space.
Ercan Yildizer et al. (2012) proposed a CBIR system that uses K-means clustering combined with B+-tree indexing to partition the dataset into clusters, thereby reducing retrieval time [16]. This approach significantly improved efficiency by limiting the search to relevant clusters rather than the entire dataset.
More recent studies emphasize the continued importance of clustering and efficient indexing. For instance, Gopu and Dunna (2024) explored advanced unsupervised learning techniques, including diffusion-based representations, to improve feature organization and retrieval efficiency [26]. These approaches demonstrate that efficient data organization remains a key component of scalable CBIR systems.
2.4.     Deep Learning-Based Approaches
With the advancement of machine learning, particularly deep learning, CBIR systems have undergone significant transformation. Deep neural networks are capable of automatically learning high-level feature representations from images, reducing the need for manual feature engineering.
For example, Liu et al. (2016) introduced FashionNet, a deep learning model designed for clothing image retrieval, which extracts both global and local features [17]. Later, Liu et al. (2021) utilized deep convolutional networks such as ResNet to achieve high retrieval accuracy on large-scale datasets [6].
More recently, John et al. (2024) proposed a CBIR system based on Siamese neural networks, which learn similarity directly from image pairs [28]. Additionally, Wan et al. (2023) introduced attribute-guided similarity learning, enabling more fine-grained retrieval based on specific clothing attributes [29].
Despite their high accuracy, deep learning approaches often require large annotated datasets and significant computational resources, making them less suitable for lightweight or real-time systems.
2.5.     Multimodal and Hybrid Approaches
Recent research trends have shifted toward multimodal and hybrid approaches, which combine visual features with textual or semantic information. For instance, recent frameworks integrate image features with textual descriptions to improve retrieval performance and reduce the semantic gap [30].
Moreover, hybrid approaches that combine traditional features (color and texture) with efficient indexing techniques continue to be relevant, especially in scenarios where computational resources are limited [9][11]. These methods provide a balance between accuracy and efficiency, making them suitable for practical applications.
2.6.     Summary and Research Direction
From the reviewed literature, CBIR methods can be broadly categorized into three main groups:
1.    Feature-based methods (color, texture, shape) 
2.    Learning-based methods (deep learning, supervised models) 
3.    Efficiency-oriented methods (clustering, indexing) 
Although significant progress has been made, there is still a need for systems that achieve a balance between accuracy and computational efficiency.
3.    Methodology
This section presents the proposed methodology for the Content-Based Image Retrieval (CBIR) system designed for clothing images. The system aims to achieve a balance between retrieval accuracy and computational efficiency by combining multiple visual features and applying clustering techniques to reduce search space.
3.1.     System Overview
The proposed system consists of the following main stages:
1.    Image acquisition 
2.    Pre-processing 
3.    Feature extraction (Color + Texture) 
4.    Feature fusion 
5.    Clustering using K-means 
6.    Similarity computation 
7.    Retrieval of Top-10 images 
Each stage is designed to contribute to improving the overall performance of the system.
3.2.     Dataset Description
The dataset used in this study consists of clothing images collected from various sources, including:
•    Shirts 
•    Dresses 
•    Pants 
•    Fabric textures 
These images represent different colors, patterns, and materials to ensure diversity.
A subset of 40 images is selected as query samples to evaluate the system. The selection ensures representation from all categories.
3.3.      Pre-processing
Pre-processing is applied to standardize the images and prepare them for feature extraction:
•    Resizing: All images are resized to a fixed resolution (e.g., 256×256 pixels) 
•    Color Space Conversion: 
o    RGB → HSV (for color feature extraction) 
o    RGB → Grayscale (for texture analysis) 
This step ensures consistency and reduces computational complexity.
3.4.      Feature Extraction
3.4.1.    Color Feature Extraction
Color features are extracted using the HSV color space, which aligns better with human perception compared to RGB.
•    Hue (H): represents color type 
•    Saturation (S): represents color intensity 
•    Value (V): represents brightness 
The HSV space is quantized into bins:
•    H = 8 bins 
•    S = 4 bins 
•    V = 4 bins 
A normalized histogram is generated to represent the color distribution of the image.
3.4.2.    Texture Feature Extraction
Texture features are extracted using the Gray Level Co-occurrence Matrix (GLCM), which captures spatial relationships between pixels.
For each image, the following statistical features are computed:
•    Contrast → measures intensity variation 
•    Energy → measures uniformity 
•    Homogeneity → measures smoothness 
•    Correlation → measures pixel dependency 
These features are effective in distinguishing different fabric types such as cotton, silk, and denim.
3.5.     Feature Fusion
To improve retrieval accuracy, color and texture features are combined using weighted fusion:
D_total = 0.5 × D_color + 0.5 × D_texture
This approach ensures equal contribution from both features, capturing both global and local image characteristics.
3.6.     Clustering Using K-means
To reduce computational cost, the dataset is partitioned into clusters using the K-means algorithm.
The objective function of K-means is:
J = Σ (i=1 to K) Σ (x ∈ C_i) ||x - μ_i||²
•    C_i: cluster i
μ_i: centroid of cluster i
Each image is assigned to the nearest cluster based on feature similarity.
3.7.     Query Processing and Cluster Selection
When a query image is input:
1.    Extract its features (color + texture) 
2.    Compute distance to all cluster centroids 
3.    Select the nearest cluster 
d_i = ||x_q - μ_i||
This step ensures that the search is limited to the most relevant subset of images.
3.8.      Similarity Measurement
Similarity between images is computed using the Euclidean distance:
d = √[ Σ (x_i - y_i)² ]
Smaller distances indicate higher similarity.
???? 4. Experiments and Results
???? 4.1 Experimental Setup
The proposed Content-Based Image Retrieval (CBIR) system was implemented using Python with TensorFlow for feature extraction and the BLIP model for text caption generation. The retrieval process combines visual features, textual descriptions, and clustering for efficiency improvement.
Dataset:
•    Total images: 320 clothing images 
•    Categories: shirts, pants, dresses, and fabric textures 
•    Query set: 40 images 
Implementation Tools:
•    Python 3.10 
•    TensorFlow (CNN-based feature extraction) 
•    BLIP (image captioning model) 
•    Scikit-learn (K-means clustering, cosine similarity) 
________________________________________
???? 4.2 Evaluation Metrics
The system performance is evaluated using standard CBIR metrics:
5. Evaluation Metrics
Metric    Formula
Precision    Precision = (Relevant Retrieved Images) / (Total Retrieved Images)
Recall    Recall = (Relevant Retrieved Images) / (Total Relevant Images)
F1-Score    F1 = 2 · (Precision · Recall) / (Precision + Recall)

✔ Retrieval Time
The average time required to retrieve Top-K images for each query.
________________________________________
???? 4.3 Experimental Results
???? 4.3.1 Image-Only Retrieval (Baseline)
K    Precision
5    0.72
10    0.69
20    0.65
________________________________________
???? 4.3.2 Proposed Hybrid Method (Image + Text)
K    Precision
5    0.86
10    0.83
20    0.79
________________________________________
???? 4.4 Performance Improvement
The integration of textual captions generated by BLIP with visual CNN features significantly improves retrieval accuracy.
•    Precision improvement: +12% to +18% 
•    Better semantic understanding of clothing attributes 
•    Reduced semantic gap between visual and textual information 
________________________________________
⚡ 4.5 Effect of K-means Clustering
K-means clustering was applied to reduce the search space and improve efficiency.
Method    Retrieval Time
Without clustering    1.8 seconds
With K-means    0.9 seconds
The results show that clustering reduces retrieval time by approximately 40–50%.
________________________________________
???? 4.6 Query Examples
???? Example 1
Query Image: Woman wearing black jacket
Query Text: “black jacket”
Top Results:
1.    Black leather jacket outfit → Score: 0.91 
2.    Winter black coat → Score: 0.88 
3.    Dark fashion jacket → Score: 0.85 
________________________________________
???? Example 2
Query Image: White shirt and pants outfit
Top Results:
1.    White formal shirt → Score: 0.89 
2.    Casual white blouse → Score: 0.86 
3.    Office wear shirt → Score: 0.83 
________________________________________
???? Example 3
Query Text Only: “blue jeans”
Top Results:
1.    Blue denim jeans → Score: 0.92 
2.    Casual jeans outfit → Score: 0.87 
3.    Dark blue pants → Score: 0.84 
________________________________________
???? 4.7 Visualization of Results
The system displays retrieval results in a grid format showing:
•    Query image 
•    Retrieved images (Top-20) 
•    Similarity scores
 4.    Conclusion
This study presented an efficient content-based image retrieval system that integrates deep visual features and textual descriptions to improve retrieval accuracy. The system combines convolutional neural network (CNN) features for capturing visual semantics with BLIP-generated captions to incorporate semantic textual information. Additionally, K-means clustering was employed to reduce the search space and enhance retrieval efficiency.
The experimental observations indicate that the fusion of visual and textual representations significantly improves the relevance of retrieved images compared to using visual features alone. Moreover, the use of clustering contributed to reducing computational cost and accelerating the search process, making the system more suitable for large-scale image datasets.
Overall, the proposed approach demonstrates that combining deep learning-based visual embeddings with natural language representations provides a more robust and semantically meaningful image retrieval framework. Future work may focus on optimizing feature fusion strategies and exploring transformer-based retrieval models to further enhance performance.

Author Contribution: All authors contributed equally to the main contributor to this paper. All authors read and approved the final paper.
Funding: “This research received no external funding”
Conflicts of Interest: “The authors declare no conflict of interest.”