To achieve rapid and non-destructive identification of early bruises in Actinidia arguta, this study proposed a parallel modeling strategy based on spectra and Gramian angular field (GAF) images. Hyperspectral data were collected at different times during the early stage of bruising, and the raw spectra were denoised using adaptive wavelet transform (AWT). Based on the denoised one-dimensional spectral data, four classification models were developed to classify the occurrence time of bruising in A. arguta fruits: support vector machine (SVM), one-dimensional convolutional neural network (1D-CNN), long short-term memory network (LSTM), and stacking ensemble learning. Meanwhile, the study innovatively introduced the GAF method to transform the spectra into two-dimensional images, which were used in combination with the YOLOv8n model for the classification of bruising time. The results demonstrated that the models differed in classification performance and computational efficiency. Among them, the stacking model, which leverages the complementary strengths of multiple base learners, achieved the best classification performance, with accuracy, recall, precision and F1-score all reaching over 97.78%, and a running time of 49 seconds. The YOLOv8n model exhibited comparable accuracy, recall, precision, and F1-score to the stacking model, and although it incurred a higher computational cost (running time of 25 min and 22 s), its end-to-end image modeling capability and superior feature visualization significantly enhanced the interpretability of the model. The dual-path modeling strategy proposed in this study provides a new approach for the rapid and accurate detection of early bruises in A. arguta fruits, offering theoretical support and guidance for the development of intelligent postharvest grading and automated sorting systems.

In order to quickly and effectively identify the slight damage of white button mushroom (Agaricus bisporus), the near-infrared hyperspectral images (400‒1000 nm) of the intact mushroom and the mushroom with different degrees of damage caused by vibration for 60 and 120 s were recorded. It was found that the spectra of the three types of samples in the wavelength range 450‒750 nm were obviously different. Compared to standard normal variable transformation and multivariate scattering correction, Savitzky Golay (SG) smoothing was determined as a better pretreatment method. The successive projection algorithm (SPA) and competitive adaptive reweighted sampling (CARS) were applied to extract the vibration-specific wavelengths. Based on the gray-level co-occurrence matrix, the texture features of the region of interest of the feature image at 500 nm were extracted. A model to discriminate the degree of damage of mushrooms was established using partial least square discriminant analysis (PLS-DA), BP neural network and extreme learning machine (ELM) based on the spectral information and the texture feature information, separately. The results showed that the PLS-DA model with each of the two feature sets as input had better performance than the other two models. With the spectral information as input, the average recognition accuracies for the training set and the test set were 93.33% and 91.11%, while those with the texture feature information as input were 88.89% and 86.67%, respectively. Finally, a PLS-DA classification model was established based on the spectrum-texture fusion information, whose overall classification accuracies for the training set and the test set were 97.78% and 95.56%, respectively. The predictive performance of the model based on the spectrum-texture fusion information was better than the model based on the single information. Therefore, the application of spectrum-texture fusion information-based modeling can improve the detection accuracy of white button mushrooms with different degrees of damage, which provides theoretical support for the storage and classification of white button mushrooms.

In this study, a custom apparatus was utilized to obtain blueberry samples with high and low levels of mechanical injury (DI and DII group), which were subsequently stored at 4 or 25 ℃ and analyzed using low-field nuclear magnetic resonance (LF-NMR) spectroscopy. Spectroscopic data and proton density images were collected at 0, 6, and 24 h postdamage. The results demonstrated a significant impact of injury on the water state and distribution of blueberries. Notably, the cell wall water content A₂₁ decreased, whereas the cytoplasm water content A₂₂ increased over time post-damage. The A₂₂ of the high injury group increased by 1.45 times at 24 h compared with that at 0 h, while the vacuole water content A₂₃ and the total water content A₂ notably declined, indicating a rapid redistribution and transfer of water from damaged blueberries to the external environment. As the degree of damage increased, the major peak of vacuole water shifted to the right and decreased significantly. Furthermore, increasing storage temperature further exacerbated these changes. At 25 ℃, the emergence of 4 relaxation peaks at both 6 and 24 h after damage indicated a significant deterioration in the quality of blueberries. Compared with 25 ℃, storage at 4 ℃ was more conducive to maintaining fruit integrity and slowing down the rate of moisture loss from damaged blueberries. The experimental results provide a theoretical basis and data reference for the rapid detection and mechanistic understanding of mechanical damage in blueberries.

To rapidly and accurately identify blueberry fruit maturity under complex field conditions, a lightweight detection model for blueberry maturity (MSC-YOLOv8)was proposed based on YOLOv8n. YOLOv8n, a recent addition to the YOLO series, offers the fastest detection speed and the lightest architecture, albeit with slightly lower recognition accuracy. Considering these attributes, YOLOv8n was chosen as the foundational model for blueberry maturity recognition. Compared to traditional object detection algorithms, YOLOv8 boasts a significantly faster detection speed. By using a single forward propagation to obtain detection results for all targets simultaneously, YOLOv8 achieves low latency, making it suitable for real-time applications. Additionally, YOLOv8 demonstrates high detection accuracy, rivaling advanced object detection algorithms, and excels particularly in detecting small targets. These advantages make YOLOv8 a superior choice for this application. Before training the model, the original blueberry datasets need to be augmented using Deep Convolution Generative Adversarial Networks (DCGAN) to generate additional blueberry images.This aims to generate diverse blueberry images to enhance the datasets, thereby improving the model's capability for detecting and recognizing blueberries in complex field environments. Firstly, to reduce the number of parameters and improve the model's running speed, MobileNetV3 was employed as the backbone feature extraction network, which was conducive to detection in complex field environments. Secondly, to enhance the network's ability to extract features from blueberries, the Convolutional Block Attention Module (CBAM) was integrated into the backbone feature extraction network. Finally, by introducing SCYLLA-IoU (SIoU) as the bounding box regression loss function for YOLOv8n, the issue of angle mismatch between the ground truth and predicted boxes was addressed, further improving the accuracy of blueberry maturity identification. In terms of the backbone feature extraction network, compared with MobileNetV3, ShuffleNet, VanillaNet, and YOLOv8's own backbone CspdarkNet, the experimental results showed that MobileNetV3 as the backbone network of YOLOv8 had the highest mAP value. Additionally, CBAM demonstrated the best mAP performance in various attention mechanisms such as CBAM, SE, ShuffleAttention and MHSA. When assessing bounding box regression loss functions, SIoU outperformed GIoU, EIoU, and CIoU in terms of mAP. Through three key improvements, this study proposed a lightweight network model, MSC-YOLOv8, for blueberry fruit maturity detection,ensuring the real-time performance and accuracy. The results showed that the improved MSC-YOLOv8 model was 3.9 percentage points higher than that of YOLOv8. The average detection time was reduced by 3.97 ms compared with the original YOLOv8. Compared with SSD and CenterNet models, the improved model achieved better results on the blueberry dataset, and the average precision mean mAP is increased by 4.6 and 1.1 percentage points, respectively, which has advantages in detection speed and accuracy. This research provides technical support for blueberry picking robots to perform picking work under complex conditions in the fields.