Each solar cell is characterized at the end-of-line using current-voltage ( IV) measurements, except shingle cells, due to multiplied measurement efforts. Therefore, the respective host cell quality is adopted for all resulting shingles, which is sufficient for samples with laterally homogeneous quality. Yet, for heterogeneous defect distributions, this procedure leads to (i) loss of high quality shingles due to defects on neighboring host cell parts, (ii) increased mismatch losses due to inaccurate binning and (iii) lack of shingle-precise characterization. In spatially resolved host measurements, such as electroluminescence images, all shingles are visible along with their properties. Within a comprehensive experiment 840 hosts and their resulting shingles are measured. Thereafter, a deep learning model has been designed and optimized which processes host-images and determines IV parameters like efficiency or fill factor, IV curves and binning classes for each shingle cell. The efficiency can be determined with an error of 0 .06 % abs enabling a 13 % abs improvement in correct assignment of shingles to bin classes compared to industry standard. This results in lower mismatch losses and higher output power on module level as demonstrated within simulations. Also IV curves of defective and defect-free shingle cells can be derived with good agreement to actual shingle measurements.