A multi-task deep learning model based on transformer for simultaneously evaluating the TVB-N and TVC contents of chicken breasts using two different hyperspectral imaging

Accurate assessment of freshness is crucial for ensuring quality and safety in the chicken meat industry. This study developed a Multi-task Interleaved Group Transformer Model (MIGTM) integrating dual hyperspectral imaging (HSI) data to simultaneously predict the total volatile basic nitrogen (TVB-N) and total viable count (TVC) in chicken breasts. The MIGTM demonstrated excellent predictive performance with $R_V^2$ of 0.9040 and 0.9499 for TVB-N and TVC, respectively, representing improvements of 4.48 % and 1.61 % over optimized chemometric models. Compared with single-task models, the MIGTM exhibited improvements of 1.84 % and 1.40 % for TVB-N and TVC prediction, respectively, while reducing computational cost by 50 %. The MIGTM outperformed existing CNN- and Transformer-based models in accuracy and stability by effectively leveraging complementary information from dual-spectral sources. The MIGTM combined with HSI provides a reliable, nondestructive solution for batch-level chicken freshness detection, offering significant potential for industrial implementation in meat quality assessment.