Visual search is transforming product discovery in e-commerce by allowing users to find similar or identical items using images rather than relying on textual input. Traditional keyword-based search methods are often limited by the ambiguity and variability of user queries, especially when users cannot accurately describe the product they are seeking. To address this limitation, we propose a deep computer vision-based visual search engine specifically designed for e-commerce platforms. Our system combines the strengths of convolutional neural networks (CNNs) for robust feature extraction with scalable approximate nearest neighbor (ANN) algorithms to enable real-time image similarity retrieval across large product catalogs. To further refine the accuracy of visual matching, we integrate deep metric learning using a Siamese network trained with triplet loss, which effectively captures fine-grained visual distinctions and enhances embedding space separation.
We evaluate our framework on two diverse datasets: DeepFashion, consisting of over 800,000 fashion images, and E-Shop Electronics, comprising 200,000 consumer electronics product images. Experimental results show that our approach achieves significant improvements in Precision @ 5, Recall @ 10, and mean average precision (mAP) over baseline models using traditional CNN embeddings and Euclidean distance. Furthermore, user studies demonstrate a 15% increase in conversion rates and a 27% rise in average session duration when using the visual search interface compared to conventional keyword-based systems. These results confirm that our visual search engine not only delivers superior retrieval performance but also enhances user engagement and satisfaction. This work establishes a scalable and intelligent framework for visually driven product discovery in modern e-commerce environments.
International Journal of Computing and Artificial Intelligence
Research Journals