Design and Application of a Custom Late Fusion Layer for Image-Numerical Milk Quality Analysis

This paper presents a custom late fusion multimodal deep learning technique for milk quality classification by integrating visual and numerical features. Top-performing unimodal models such as MobileNet, Inception V3, and DenseNet for visual data, and LightGBM, CatBoost, and XGBoost for numerical data were identified through comparative evaluation. The proposed concatenation-with-proposed-layers fusion model achieved a peak testing accuracy of 99.77%, matching or surpassing alternative fusion techniques while employing fewer layers for improved computational efficiency. Comparative experiments demonstrated superior performance over max pooling, majority voting, and weighted average methods, with notable robustness across nine visual–numerical model pairings. A human-centered study further validated the approach, showing that combining visual and numerical inputs improved classification accuracy by up to 45.1% in certain cases. The results highlight the proposed model’s effectiveness, stability, and applicability in quality control and safety-critical domains, with potential extension to other multimodal classification tasks requiring high precision.