Since impurities produced during walnut processing can cause serious harm to human health, strict quality control must be carried out during production. However, most detection equipment still uses photoelectric detection technology to automatically sort heterochromatic particles, which is unsuitable for detecting endogenous foreign bodies with similar colors. Therefore, this paper proposes an improved YOLOv4 deep learning object detection algorithm, WT-YOLOM, for detecting endogenous impurities in walnuts—namely, oily kernels, black spot kernels, withered kernels, and ground nutshells. In the backbone of the model, a lightweight MobileNet module was used as the encoder for the extraction of features. The spatial pyramid pooling (SPP) structure was improved to spatial pyramid pooling—fast (SPPF), and the model size was further reduced. Loss function was replaced in this model with a more comprehensive SIoU loss. In addition, efficient channel attention (ECA) mechanisms were applied after the backbone feature map to improve the model’s recognition accuracy. This paper compares the recognition speed and accuracy of the WT-YOLOM algorithm with the Faster R-CNN, EfficientDet, CenterNet, and YOLOv4 algorithms. The results showed that the average precision of this model for different kinds of endogenous impurities in walnuts reached 94.4%. Compared with the original model, the size was reduced by 88.6%, and the recognition speed reached 60.1 FPS, which was an increase of 29.0%. The metrics of the WT-YOLOM model were significantly better than those of comparative models and can significantly improve the detection efficiency of endogenous foreign bodies in walnuts.

To address the complexity of crispness evolution in Lin'an mountain walnuts under eight distinct thermal processing methods, this study proposes a novel Multi-modal Cross-Attention Fusion Network (MCAFNet), aiming to achieve intelligent and non-destructive evaluation of the physical quality of nut-based foods. First, the Markov Transition Field (MTF) is creatively introduced to transform one-dimensional electronic nose time-series signals into multi-channel images, combined with multi-scale parallel convolutions to capture implicit temporal dynamic features across different time scales. Second, spectral indices and a reconstruction-stacking strategy are adopted to synergistically extract local and global features via multi-stage convolutions, significantly enhancing the representation capability of spectral information. Third, to overcome the semantic gap between multi-modal data, MCAFNet designs a Dual-Branch Feature Fusion (DBFF) module to perform feature-level fusion on one-dimensional image and spectral features, while introducing a bidirectional cross-attention mechanism to achieve deep interaction and model-level fusion between the spatial features of pseudo-spectral images and the temporal features of MTF images. Finally, the prediction results are output via a hybrid stacking ensemble strategy. Experimental results demonstrate the superior performance of MCAFNet, achieving a coefficient of determination (R2) of 0.968 and a residual predictive deviation (RPD) of 5.578 for crispness prediction. In conclusion, MCAFNet not only offers an efficient solution for the non-destructive crispness evaluation of nut-based products but also establishes a solid theoretical and practical foundation for intelligent quality monitoring in food engineering.