Content-Based Image Retrieval (CBIR) is a fundamental task in computer vision, aiming to retrieve visually similar images given a query. Traditional CBIR methods depend solely on the query’s feature representation, which may be suboptimal due to variations in viewpoint, lighting, or occlusion. Although Convolutional Neural Networks (CNNs) have improved feature extraction, they remain sensitive to spatial transformations and structural changes and often struggle to capture high-level semantic relationships. As a result, their discriminative power is limited, especially when the query image differs significantly from the target images. In this study, we propose a retrieval method that refines feature representations by leveraging information from k-nearest neighbors of the query image. Instead of using only the original query feature, we construct a new feature vector through a weighted aggregation of its neighbors’ features. This strategy enhances the robustness and relevance of the representation. The proposed method is evaluated on some benchmark datasets using common CNN backbones, including VGG16, ResNet50, EfficientNetB0, and DenseNet201. Experimental results demonstrate that our approach consistently improves retrieval accuracy over conventional methods that rely solely on the query’s original features.