<p>Sketch-Based Image Retrieval (SBIR) enables searching large image databases using freehand sketches, but remains challenging due to the substantial domain gap between sparse drawings and detailed photos. Handcrafted features and deep learning methods have made progress, yet most operate only in the spatial domain, overlooking frequency cues that capture structural information. We propose LTL-SBIR, a deep Siamese framework that integrates a learnable frequency-domain transform layer. Unlike fixed Fourier or wavelet transforms, this layer adaptively learns discriminative bases through end-to-end training, enhancing cross-domain matching between sketches and photos. The model captures multi-scale structural cues while suppressing noise and style variation. Experiments on three fine-grained benchmarks QMUL-Shoe-V2, QMUL-Chair, and Sketchy (Extended) demonstrate consistent improvements over state-of-the-art baselines. LTL-SBIR achieves mAP scores of 64.8, 62.3, and 59.9%, respectively, with notable gains in P@10 and P@100, confirming robust retrieval at both fine-grained and large-scale levels. An ablation study highlights the key role of the frequency-domain layer, while efficiency analysis shows that accuracy gains incur only modest additional FLOPs and latency. These results establish frequency-aware learning as an effective approach for fine-grained SBIR. Beyond accuracy, the framework demonstrates deployment readiness, achieving a balance between performance and computational efficiency. Code and experimental configurations are available at: <a href="https://github.com/mohammedalmohmdy/ltl-sbir-pytorch.git">https://github.com/mohammedalmohmdy/ltl-sbir-pytorch.git</a></p>