Industrial defect detection is crucial for ensuring product quality and production efficiency, playing a pivotal role in advancing smart manufacturing. This paper reviews defect detection technologies for various industrial products, including metals, textiles, and printed circuit boards, and introduces an innovative classification system. It also offers a detailed analysis of recent developments and practical applications of large models in industry defect detection. First, the basic principles of industrial defect detection are outlined. The detection methods are then categorized into three main groups: traditional image processing, machine learning, and deep learning, with their principles, case studies, limitations, and future development directions analyzed. Traditional methods consist of image preprocessing, segmentation, and feature extraction. Machine learning methods are divided into point-distance-based, hyperplane-based, tree-based, and neural network-based classification algorithms. Deep learning models are classified into two types: accuracy-oriented and efficiency-oriented. The paper organizes industrial defect datasets by type (multi-product and single-product), evaluates data quality and availability, and summarizes common evaluation metrics for accuracy, efficiency by task requirements. It also compares the latest methods on two public datasets to guide further research in defect detection. Real-world examples illustrate the end-to-end process, from data processing and hardware configuration to model training and deployment, while exploring the value and limitations of these technologies from the perspective of industry stakeholders. Finally, a systematic analysis of the key challenges and corresponding solutions is presented at the data and performance levels, and looks forward to the future direction of technological development, highlighting innovative paths and application potentials.

High-utility itemset mining (HUIM) is an important research topic in the data mining field. Typically, traditional HUIM algorithms must handle the exponential problem of huge search space when the database size or number of distinct items is very large. As an alternative and effective approach, evolutionary computation (EC)-based algorithms have been proposed to solve HUIM problems because they can obtain a set of nearly optimal solutions in limited time. However, it is still time-consuming for EC-based algorithms to find complete high-utility itemsets (HUIs) in transactional databases. To address this problem, we propose an HUIM algorithm based on an improved genetic algorithm (HUIM-IGA). In addition, a neighborhood exploration strategy is proposed to improve search efficiency for HUIs. To reduce missing HUIs, a population diversity maintenance strategy is employed in the proposed HUIM-IGA. An individual repair method is also introduced to reduce invalid combinations for discovering HUIs. In addition, an elite strategy is employed to prevent the loss of HUIs. Experimental results obtained on a set of real-world datasets demonstrate that the proposed algorithm can find complete HUIs in terms of the given minimum utility threshold, and the time-consuming of HUIM-IGA is relatively lower when mining the same number of HUIs than state-of-the-art EC-based HUIM algorithms.

The complexity of urban road scenes at night and the inadequacy of visible light imaging in such conditions pose significant challenges. To address the issues of insufficient color information, texture detail, and low spatial resolution in infrared imagery, we propose an enhanced infrared detection model called LFIR-YOLO, which is built upon the YOLOv8 architecture. The primary goal is to improve the accuracy of infrared target detection in nighttime traffic scenarios while meeting practical deployment requirements. First, to address challenges such as limited contrast and occlusion noise in infrared images, the C2f module in the high-level backbone network is augmented with a Dilation-wise Residual (DWR) module, incorporating multi-scale infrared contextual information to enhance feature extraction capabilities. Secondly, at the neck of the network, a Content-guided Attention (CGA) mechanism is applied to fuse features and re-modulate both initial and advanced features, catering to the low signal-to-noise ratio and sparse detail features characteristic of infrared images. Third, a shared convolution strategy is employed in the detection head, replacing the decoupled head strategy and utilizing shared Detail Enhancement Convolution (DEConv) and Group Norm (GN) operations to achieve lightweight yet precise improvements. Finally, loss functions, PIoU v2 and Adaptive Threshold Focal Loss (ATFL), are integrated into the model to better decouple infrared targets from the background and to enhance convergence speed. The experimental results on the FLIR and multispectral datasets show that the proposed LFIR-YOLO model achieves an improvement in detection accuracy of 4.3% and 2.6%, respectively, compared to the YOLOv8 model. Furthermore, the model demonstrates a reduction in parameters and computational complexity by 15.5% and 34%, respectively, enhancing its suitability for real-time deployment on resource-constrained edge devices.

The accurate detection of occluded tomatoes in complex greenhouse environments remains challenging due to the limited feature representation ability and high computational costs of existing models. This study proposes SWMD-YOLO, a lightweight multi-scale detection network optimized for greenhouse scenarios. The model integrates switchable atrous convolution (SAConv) and wavelet transform convolution (WTConv) for the dynamic adjustment of receptive fields for occlusion-adaptive feature extraction and to decompose features into multi-frequency sub-bands, respectively, thus preserving critical edge details of obscured targets. Traditional down-sampling is replaced with a dynamic sample (DySample) operator to minimize information loss during resolution transitions, while a multi-scale convolutional attention (MSCA) mechanism prioritizes discriminative regions under varying illumination. Additionally, we introduce Focaler-IoU, a novel loss function that addresses sample imbalance by dynamically re-weighting gradients for partially occluded and multi-scale targets. Experiments on greenhouse tomato data sets demonstrate that SWMD-YOLO achieves 93.47% mAP50 with a detection speed of 75.68 FPS, outperforming baseline models in accuracy while reducing parameters by 18.9%. Cross-data set validation confirms the model’s robustness to complex backgrounds and lighting variations. Overall, the proposed model provides a computationally efficient solution for real-time crop monitoring in resource-constrained precision agriculture systems.

The fatigue driving detection process faces issues such as a large number of parameters, low accuracy and insufficient continuous detection. To address these, this paper proposes a method combining enhanced YOLOv5s and DeepSort for fatigue driving detection. First, the improved Mobilenet_ECA lightweight backbone is introduced to reconstruct the YOLOv5s backbone part. Then, the CSPDarknet53(C3) module in the neck network is integrated with the Triplet Attention Module (TAM) to enhance the fusion of contextual information and improve the network’s facial feature extraction capability. In addition, FocalEIOU Loss is used to optimize the problem of large errors in Complete Intersection over Union (CIoU) Loss prediction box regression calculations. Next, the DeepSort facial feature tracking algorithm is combined for continuous classification tracking to optimize the problem of driver facial information loss. Finally, the facial feature detection results are combined with the number of consecutive eyeclosed frames, the number of consecutive yawning frames, and the Percentage of Eyelid Closure over the Pupil over Time (PERCLOS) score to build a fatigue driving detection model. The experimental results show that compared with the baseline model YOLOv5s, the algorithm proposed in this article has improved mAP@0.5 and P by 1% and 1.8%, Params decreased by 56.3%, FLOPs decreased by 63.2%, and the model size is only 6.4 MB. The final recognition rate of fatigue and nonfatigue driving was as high as 97.4%. It is verified that the algorithm in this paper can maintain high detection accuracy while being lightweight, and can effectively identify driver states, providing strong support for the deployment of vehicle edge devices.

Hyperspectral images (HSIs) are highly complex, containing an enhanced spectral dimension compared to conventional images. Deep learning methods are increasingly being applied to process this three-dimensional data for hyperspectral image classification (HSIC). Existing convolutional and transformer-based methods often struggle with capturing fine-grained spectral-spatial dependencies, achieving high accuracy while balancing computational complexity. In order to solve these problems, we propose a novel SXSFormer network for HSIC. This novel approach integrates Squeeze and Expansion (SX) blocks into the Swin Transformer architecture to enhance feature extraction and attention mechanisms in the model. At its core, the novel SX Block recalibrates channel-wise features by temporarily expanding and then compressing channel dimensionality, allowing the model to focus on informative spectral bands and capture complex interdependencies. The SXSFormer, equipped with window-based multi-head self-attention, efficiently captures long-range dependencies with reduced computational complexity by partitioning input data into non-overlapping windows. Additionally, the SX block’s integration within the Swin Transformer enables improved global contextual understanding by selectively weighting the feature maps. We further refine the architecture using various components, such as patch extraction and embedding layers, and a patch merging strategy, ensuring efficient multi-scale feature extraction. Extensive experiments on four benchmark HSI datasets (SA, IP, PU, and KSC) demonstrate that the proposed model achieves remarkable test accuracies of 99.97%, 98.15%, 99.63%, and 98.14%, respectively, outperforming existing state-of-the-art methods. Our proposed approach also shows good generalization ability when applied to new datasets. Overall, our proposed approach represents a promising direction for HSIC.