<p>This paper introduces DSFLNet, a novel dual-stream feature learning network for content-based video retrieval. The model addresses key challenges in video retrieval by integrating EfficientNet for spatial feature extraction and an Enhanced Transformer with GRU to model temporal dependencies. EfficientNet, with its residual connections, ensures efficient extraction of high-quality spatial features, while the Enhanced Transformer with GRU captures complex temporal dynamics over the sequence of frames. A unique attention-based feature fusion module, combining both channel-wise and pixel-wise attention, is proposed to effectively integrate the spatial and temporal features, enabling more accurate retrieval. Additionally, the model employs Contrast Limited Adaptive Histogram Equalization preprocessing to enhance frame quality, making it more resilient to noisy and inconsistent video data. The proposed DSFLNet significantly improves retrieval performance by enhancing both the accuracy and efficiency of the process. This approach has broad applicability across various real-world domains, including medical imaging, video surveillance, and e-commerce, where large-scale, precise video retrieval is crucial.</p> Graphical abstract <p></p>