A Large-Scale Dataset and Robust Multifeature Representation With Maximum Correlation-Based Feature Fusion and Matching for Apparel Image Retrieval

Abstract

Finding the correct match to a probe image from a vast amount of data is critical for the online retrieval of apparel images. These images are captured under an uncontrolled environment (e.g., viewpoint and illumination changes); therefore, such type of data is extremely challenging in Content-Based Image Retrieval (CBIR) research. Even in Google searches, most of the time the query results are provided with inaccurate results or duplicate results due to the minor variations between apparel. Another major challenge is that the extracted feature vector dimensions are too high and difficult to handle. In this paper, a method named Multifeature Representation with Maximum Correlation-based Feature Fusion, and Matching (MFR-MCF²M) is proposed for apparel retrieval. This method consists of three modules: (1) Multifeature Representation Module (MFR-M), (2) Maximum Correlation-based Feature Fusion Module (MCF²-M) and (3) Multifeature Matching Module (MFM-M). In the MFR module, the shape, texture and deep features of apparel images are extracted using a Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP) and a pretrained deep CNN model, respectively. Also, the dimensionality of extracted features is reduced using the proposed Feature Subselection (FSS) method. The MCF module is implemented to measure the maximum correlation between reduced feature vectors. Finally, MCF² is performed using Euclidean distance and a generated Feature Correlation Vector (FCV) to improve the retrieval accuracy and as the benchmark to assess the efficacy of the proposed method. In addition, a new large-scale dataset named Apparel Images Gallery (AIG), which consists of 130,000 images, has been provided to the community. The performance of the proposed MFR-MCF²M method is evaluated on three datasets, including two publicly available datasets and the proposed AIG dataset. The retrieval results are obtained after passing through the threshold function of both the Euclidean distance and the computed FCV. The proposed method achieved an accuracy of 78.3% on the clothing dataset, 94.8% on the CR dataset and 89.1% on the proposed AIG dataset. Consequently, the MFR-MCF²M outperformed state-of-the-art (SOTA) apparel retrieval methods.