With the rapidly growing number of digital photos being taken using different devices in recent years, significant attention has been brought to improving the ability to match these images. However, the reliance on traditional Content-Based Image Retrieval (CBIR) techniques on certain features, e.g., objects, low-level features, or colors, in these images has caused a semantic gap in the matching results. Recent techniques rely on multiple features to reduce this gap and employ artificial neural networks (ANNs) to produce a single similarity measure that represents the overall similarity between two images. Additionally, several studies have suggested that these networks better detect low-level features when processing the input image in grayscale rather than separate color channels. In this study, we propose a new methodology that allows ANNs to process colored and grayscale versions of images simultaneously, producing a more accurate similarity measure by accurately considering the high-level, low-level, and colors in the input images. The model implementation is based on the Yolo V8 neural network architecture. It is evaluated against recent state-of-the-art methods using several datasets, including MIRFLICKR-25K, NUS-WIDE, MS-COCO, Pascal VOC2007, and Pascal VOC2012. We assess the model's performance using three well-established metrics: NDCG, ACG, and wMAP. The proposed technique outperformed all existing methods in terms of NDCG and wMAP. Experimental results demonstrate that this method has also achieved high-performance measures with significant improvements and more stable results at different datasets of different images and classes, especially when the quality of the results is measured using the NDCG. Such an improvement illustrates the importance of using the grayscale version of the image as an input to the neural network to improve its ability to recognize local features better than only providing the image in RGB.
Ahmed et al. (Thu,) studied this question.