Hao Li

Eggshell cracks are a critical factor affecting egg quality and food safety, with traditional detection methods often struggling to detect fine cracks, especially under multi-colored shells and complex backgrounds. To address this issue, we propose a non-destructive detection approach based on an enhanced three-dimensional convolutional neural network (3D-CNN), named 3D-CrackNet, integrated with hyperspectral imaging (HSI) for high-precision identification and localization of eggshell cracks. Operating within the 1000–2500 nm spectral range, the proposed framework employs spectral preprocessing and optimal band selection to improve discriminative feature representation. A residual learning module is incorporated to mitigate gradient degradation during deep joint spectral-spatial feature extraction, while a parameter-free SimAM attention mechanism adaptively enhances crack-related regions and suppresses background interference. This architecture enables the network to effectively capture both fine-grained spatial textures and contiguous spectral patterns associated with cracks. Experiments on a self-constructed dataset of 400 egg samples show that 3D-CrackNet achieves an F1-score of 75.49% and an Intersection over Union (IoU) of 60.62%, significantly outperforming conventional 1D-CNN and 2D-CNN models. These findings validate that 3D-CrackNet offers a robust, non-destructive, and efficient solution for accurately detecting and localizing subtle eggshell cracks, demonstrating strong potential for intelligent online egg quality grading and micro-defect monitoring in industrial applications.

Egg variety and storage duration are key determinants of nutritional value, market pricing, and food safety. The similar external appearance of different varieties increases the risk of mislabeling, while inevitable quality deterioration during storage further complicates reliable assessment. These factors underscore the need for non-destructive, cross-temporal detection. However, prolonged storage induces pronounced spectral drift that degrades conventional models, limiting their effectiveness in real-world quality monitoring. To address this issue, we propose the Multi-Task Cross-Temporal Squeeze-and-Excitation Network (MT-CTSE-Net), a deep learning framework that integrates Convolutional Neural Networks (CNN), Squeeze-and-Excitation (SE) channel attention, and Transformer encoders to jointly perform egg variety identification across storage durations and storage period classification. The model extracts local spectral details, enhances channel-wise feature relevance, and captures long-range dependencies, while inter-task feature sharing improves generalization under temporal variation. Evaluated on near-infrared (1000–2500 nm) spectra from three commercial egg varieties (Enshi selenium-enriched, Mulanhu multigrain, Zhengda lutein), MT-CTSE-Net achieved approximately 86% accuracy (F1-score: 86.1%) in cross-temporal variety classification and about 84.2–86.4% in storage-period prediction—surpassing single-task and benchmark multi-task models. These results demonstrate that MT-CTSE-Net effectively mitigates storage-induced spectral drift and provides a robust pathway for non-destructive quality assessment and temporal monitoring in agri-food supply chains.

With the lack of sufficient prior informathon, unsupervised hyperspectral unmixing (HU) has been a pre-processing step in the HSI processing pipeline, which can provide the types of material and corresponding abundance information of HSI, to further provide assistance for downstream higher-level semantic tasks to overcome the limitation caused by mixed pixels. However, the unmixing results obtained by current unsupervised HU methods are unstable and unprecise under the guidance of the least reconstrucion error, which have no consistency with the performance high-level tasks. To solve this proplem, this article takes the hyperspectral anomaly detection (HAD) as an entry point and proposes a novel algorithm based on deep clustering which can jointly perform HU and HAD in an end-to-end manner. A mutual feedback mechanism is formed between the upstream HU process and the downstream HAD process, and through joint optimization, both two tasks can achieve relatively good performances. However, the low demensional abundance has the limited representation, which may lead to the increase of false alarm rate. To overcome this limitation, the principal components (PCs) of HSI are fused with the abundance to enhance the representation ability. Moreover, we use the re-weighted reconstruction loss strategy to enhance the role of anomalies in the HU process. Experiments performed on several real datasets verify the rationality and superiority of proposed UADNet algorithm.

The immense representation power of deep learning frameworks has kept them in the spotlight in hyperspectral image (HSI) classification. Graph Convolutional Neural Networks (GCNs) can be used to compensate for the lack of spatial information in Convolutional Neural Networks (CNNs). However, most GCNs construct graph data structures based on pixel points, which requires the construction of neighborhood matrices on all data. Meanwhile, the setting of GCNs to construct similarity relations based on spatial structure is not fully applicable to HSIs. To make the network more compatible with HSIs, we propose a staged feature fusion model called SFFNet, a neural network framework connecting CNN and GCN models. The CNN performs the first stage of feature extraction, assisted by adding neighboring features and overcoming the defects of local convolution; then, the GCN performs the second stage for classification, and the graph data structure is constructed based on spectral similarity, optimizing the original connectivity relationships. In addition, the framework enables the batch training of the GCN by using the extracted spectral features as nodes, which greatly reduces the hardware requirements. The experimental results on three publicly available benchmark hyperspectral datasets show that our proposed framework outperforms other relevant deep learning models, with an overall classification accuracy of over 97%.

The Segment Anything Model (SAM) is a versatile image segmentation model that enables zero-shot segmentation of various objects in any image using prompts, including bounding boxes, points, texts, and more. However, studies have shown that the SAM performs poorly in agricultural tasks like crop disease segmentation and pest segmentation. To address this issue, the agricultural SAM adapter (ASA) is proposed, which incorporates agricultural domain expertise into the segmentation model through a simple but effective adapter technique. By leveraging the distinctive characteristics of agricultural image segmentation and suitable user prompts, the model enables zero-shot segmentation, providing a new approach for zero-sample image segmentation in the agricultural domain. Comprehensive experiments are conducted to assess the efficacy of the ASA compared to the default SAM. The results show that the proposed model achieves significant improvements on all 12 agricultural segmentation tasks. Notably, the average Dice score improved by 41.48% on two coffee-leaf-disease segmentation tasks.

Hyperspectral unmixing (HU) based on linear mixing model (LMM) has received much attention in the remote sensing community over the past decades. However, many HU algorithms are based on the fixed endmember and do not fully exploit the inherent spatial characteristics of the hyperspectral image (HSI). In this letter, a new HU algorithm named GLR l₁ -GMM is proposed to solve these problems. We use Gaussian mixture model (GMM) to represent the endmember variability. Then, considering the local smoothness in the abundance map, a graph Laplacian regularization based on l ₁ -norm (GLR l₁ ) is embedded in the prior of abundance. Under the Bayesian framework, the objective density function leads to a maximum a posterior (MAP) problem, which can be solved by a generalized expectation-maximization (GEM) algotithm. Experiments on two real datasets demonstrates the effectiveness of proposed algorithm compared with other state-of-the-art methods. The source code is available at https://github.com/lwdinwhu/GMM-GLRl1.

Image stitching task targets to derive a large panoramic image for obtaining extensive information. However, artifacts like ghosting or geometric misalignment are inevitably generated. As a practical measure, optimal seamline detection strategies utilize the spatial information to obtain the optimal seam in RGB image stitching, but they cannot be directly used in hyperspectral image (HSI) stitching. Since the spatial information of numerous continuous bands of HSI is different, the detected seam of traditional RGB-based method in each band of HSI is divergent, which will cause visual difference and spectral distortion. To solve this problem, we propose a novel optimal seamline detection strategy via graph cuts for HSI stitching in this work. First, we employ robust feature matching and elastic warp to align multiple adjacent images into a common geometrical transformation. After that, we design a novel energy function composing both spatial and spectral information of HSI to determine an optimal seam in continuous regions with high texture consistency. Finally, we utilize the graph cuts method to eliminate visible artifacts. Our method can determine a unique optimal seam in the whole HSI for stitching so as to obtain high-quality panoramic HSI without artifacts and reduce the spectral distortion. A series of experiments verify the effectiveness and superiority of the proposed method to several advanced approaches in HSI stitching.