MNFF: Multi-nonlinear feature fusion framework for edible insect quality classification

Rapid industrialization of edible insect production and the growing demand for sustainable protein sources have increased the need for reliable quality inspection in large-scale breeding environments. However, early abnormal larvae often exhibit only weak and spatially dispersed morphological differences, making them difficult to detect using conventional vision models. To address this challenge, we propose a Multi-Nonlinear Feature Fusion (MNFF) framework for binary edible insect quality classification and instantiate it as MV-MNFF by integrating the framework with a MambaVision backbone, a convolutional block attention module (CBAM), and a classification head. Unlike conventional attention-based enhancement or simple fusion strategies that primarily reweight or aggregate features, the proposed MNFF framework explicitly models interactions among weak morphological cues by multiplicatively integrating GELU, ELU, and ReLU responses within depthwise separable and multi-scale dilated convolution stages, together with channel attention and skip connections. Experiments on mealworm larvae ( Tenebrio molitor ) and flower chafer larvae ( Protaetia brevitarsis ) showed that the proposed model improved balanced accuracy by 1.96% and 0.40%, respectively, over the baseline MambaVision, while achieving an area under the receiver operating characteristic curve (AUROC) of 1.000 on both datasets. Explainability was further examined using Grad-CAM, LIME, SHAP, and Saliency Map and quantitatively evaluated with faithfulness metrics, including average drop (%), percentage increase in confidence, positive perturbation AUC, and negative perturbation AUC. These results indicate that the proposed MNFF framework and its MambaVision-based implementation provide an effective and explainable solution for early abnormality screening and support more reliable field-level decision-making in industrial-scale edible insect farming.