Real-time and accurate quality monitoring of fruits and vegetables during cold chain logistics is of great importance for ensuring supply chain quality and reducing economic losses. However, traditional detection methods generally suffer from several core deficiencies, such as being offline, relying on unimodal information, and being unable to capture dynamic evolution. To overcome these challenges, an online freshness detection system is proposed and implemented for fruits and vegetables based on temporal multi-source information fusion. The system was designed to achieve precise online detection of fruit and vegetable freshness, providing an effective technical solution for the refined management and early spoilage warning within the cold chain supply chain, thereby significantly reducing economic losses.

A complete system was constructed, consisting of a lower-computer data acquisition node, an IoT cloud platform, and an upper-computer Qt client. The lower-computer synchronously collected environmental temporal sensing data (temperature, humidity, CO₂, ethylene) and visual temporal images of indicator tags via a self-designed portable acquisition node. A novel co-attention-based convolutional recurrent network (Co-ACRN) deep learning model was proposed for deeply mining the complex correlations between the two heterogeneous time-series data streams. This model innovatively employed a "coattention + self-attention" dual mechanism. Firstly, in the early fusion stage, a co-attention module intelligently aligned and deeply integrated visual and sensor feature sequences by constructing a cross-modal affinity matrix. Subsequently, the fused sequence was fed into a long short-term memory (LSTM) network to encode temporal cumulative effects. Finally, a self-attention module performed a global contextual review on the LSTM output to capture long-range temporal dependencies. In the specific implementation, visual features were extracted by a lightweight convolutional neural network (CNN) with two convolutional-pooling layers; the co-attention calculated weights by generating context-aware intermediate features; and the self-attention adopted the standard scaled dot-product attention mechanism. For application deployment, the model was efficiently deployed to the Qt client in the open neural network exchange (ONNX) format, achieving real-time, edge-side inference.

Experimental results showed that the proposed Co-ACRN model achieved an overall accuracy of 96.93% on the test set in the three-class mango freshness detection task, with its performance significantly surpassing that of various mainstream baselines and advanced temporal multimodal fusion models, such as modality-invariant and specific-representations for multimodal sentiment analysis (MISA), recurrent attended variation embedding network (RAVEN), multimodal transformer (MulT), and heterogeneous hierarchical message passing network (HHMPN). To verify the rationale of the model design, two sets of ablation experiments were conducted. The input-based ablation study decisively proved that the combination of "time-series information + multimodal information" is a necessary prerequisite for accurate detection, as any model relying on unimodal or static information exhibited significant performance bottlenecks. The architecture-based ablation study further confirmed the superiority of the proposed "dual-attention" system; compared to a backbone network without any attention mechanism, its accuracy was improved by more than five percentage points, and the recall rate for the critical "spoiled" category was as high as 99.16%. An in-depth analysis of the confusion matrix revealed that the vast majority of the model's errors occurred between adjacent categories with the most similar physical states, with no serious cross-category misclassifications, demonstrating its strong robustness. After being deployed on the client side, the system's single diagnosis time was less than 2 s, verifying the solution's combination of high accuracy and real-time performance.

The developed online detection system and Co-ACRN model successfully enabled the real-time, accurate, and non-destructive intelligent detection of fruit and vegetable freshness. The research findings indicate that by combining advanced co-attention and self-attention mechanisms, the fusion challenges of complex multimodal temporal data can be effectively solved. In summary, this study provides a complete solution that combines theoretical innovation with engineering practicality for the online and intelligent detection of distributed fruit and vegetable freshness, and paves new paths for the development of this field in both theory and practice.

Fresh fruits and vegetables are prone to quality deterioration during storage and transportation due to microbial proliferation and changes in enzyme activity. Although traditional quality detection methods (e.g., physicochemical analysis and microbial culture) offer high accuracy, they are destructive, time-consuming, and require expert operation, making them inadequate for the modern supply chain's demand for real-time, non-destructive detection. While advanced optical detection technologies like hyperspectral imaging provide non-destructive advantages, the equipment is expensive, bulky, and lacks portability. This study aimed to integrate fluorescence imaging technology, embedded systems, and lightweight deep learning models to develop an embedded detection system for fruit and vegetable quality deterioration, addressing the bottlenecks of high cost and insufficient portability in current technologies, and providing a low-cost, efficient solution for non-destructive quality detection of fruits and vegetables.

An embedded quality detection system based on fluorescence imaging and a ZYNQ platform was developed. The system adopted the Xilinx ZYNQ XC7Z020 heterogeneous SoC as the core controller and used 365 nm, 10 W ultraviolet LED beads as the excitation light source. A CMOS camera served as the image acquisition sensor to capture and process fluorescence images. Algorithmically, an improved, lightweight object detection model based on YOLOv8 was developed. The improved model replaced the original YOLOv8 backbone network with MobileNetV4 to reduce computational load. To further achieve lightweighting, a channel pruning technique based on the batch normalization (BN) layer's scaling factor (γ) was employed. During training, L1 regularization was applied to γ to induce sparsity, after which channels with small γ values were pruned according to a threshold (γ_threshold = 0.01), followed by fine-tuning of the pruned model. Finally, in accordance with the hardware characteristics of the ZYNQ platform, a dynamic 16-bit fixed-point quantization method was adopted to convert the model from 32-bit floating point to 16-bit fixed point, and the FPGA's parallel computing capability was utilized for hardware acceleration to improve inference speed.

Grapes and spinach were used as experimental samples in a controlled laboratory setting (26 ℃; 20%~40% humidity) over an eight-day storage experiment. Fluorescence images were collected daily, and physicochemical indices were measured simultaneously to construct ground-truth labels (spinach: chlorophyll, vitamin C; grapes: titratable acidity, total soluble solids). K-means clustering combined with principal component analysis (PCA) was used to categorize quality into three levels, "fresh" "sub-fresh" and "spoiled", based on changes in physicochemical indices, and images were labeled accordingly. In terms of system performance, the improved YOLOv8-MobileNetV4 model achieved a mean average precision (mAP) of 95.91% for the three-level quality classification. Ablation results showed that using only the MobileNetV4 backbone or applying channel pruning to the original model each reduced average detection time (by 14.0% and 29.0%, respectively) but incurred some loss of accuracy. In contrast, combining both yielded a synergistic effect: precision reached 97.04%, while recall and mAP increased to 95.24% and 95.91%, respectively. Comparative experiments indicated that the proposed model (8.98 MB parameters) outperformed other mainstream lightweight models (e.g., Faster R-CNN and YOLOv8-Ghost) in mAP and also exhibited faster detection, demonstrating an excellent balance between accuracy and efficiency.

Targeting practical needs in detecting fruit and vegetable quality deterioration, this study proposed and implemented an efficient detection system based on fluorescence imaging and an embedded platform. By integrating the MobileNetV4 backbone with the YOLOv8 detection framework and introducing BN-based channel pruning, the model achieved structured compression and accelerated inference. Experimental results showed that the YOLOv8-MobileNetV4 plus pruning model significantly reduced model size and hardware resource consumption while maintaining detection accuracy, thereby enhancing real-time responsiveness. The system's low hardware cost, compact size, and portability make it a practical solution for rapid, non-destructive, real-time quality monitoring in fruit and vegetable supply chains. Future work will focus on expanding the sample library to include more produce types and mixed deterioration levels and further optimizing the algorithm to improve robustness in complex multi-target scenarios.