| Deep learning has achieved strong performance on image data, but applying vision models to tabular numerical data remains challenging because numerical records lack spatial structure. The Numerical-to-Image Strip Encoding framework (NISE) addresses this by transforming numerical records into visual strip representations compatible with YOLO-based object detection. However, the original NISE formulation processes one record per image, causing the number of forward passes to scale linearly with the number of records. This paper introducesNISEv2, a representation-level reformulation of NISE that improves computational efficiency without modifying the YOLOv8 architecture.NISEv2 reformulates numerical classification as a high-density multi-object detection task by arranging multiple encoded records within a single 640x640 YOLO canvas. This enables multiple records to be processed in one forward pass, aligning numerical-to-image learning with YOLO's native multi-object detection paradigm. Experimental results across five benchmark datasets show that NISEv2 maintains competitive classification performance while substantially reducing computational overhead within the NISE framework. On the Adult Census Income dataset with 32,561 samples, NISEv2 achieves 0.84±0.02 accuracy, outperforming Decision Tree, SVM, and KNN, but remaining below XGBoost at 0.87±0.00. Its main advantage is computational scalability, reducing the effective number of training images per epoch from 19,536 single-record images to 127 effective training images per fold. These results indicate that high-density strip packing improves the scalability of YOLO-based numerical-to-image learning. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.