The present research suggests optimizing the Content-Based Image Retrieval (CBIR) algorithm by integrating multi-depth features from various variants of the DenseNet convolutional neural network. Traditional CBIR systems have been using low-level statistical characteristics or even single-level high-level statistical features, which constrain the retrieval accuracy. In contrast, the present work proposes a novel strategy for combining the characteristics of DenseNet-121, DenseNet-169, and DenseNet-201 models to form a unified, discriminant descriptor that encompasses both low- and high-level image features. An offline extraction of feature vectors from an image collection is stored in a feature database, and the query process is executed online using the same feature extraction pipeline. Distance measures are used to conduct similarity matching in retrieving the most relevant images. The results of the experiment demonstrate that the proposed multi-fusion of DenseNet is significantly more successful than single models, achieving a precision of 99.53 and a recall of 96.63. The model also achieves high retrieval performance in various Top-K scenarios, ensuring its robustness and scalability. Such a method offers a viable and effective solution to the problem of real-world image retrieval across various domains.