This research establishes an advanced cloud-native multimodal intelligent system architecture dedicated to automated detection and real-time monitoring of steel surface defects. The architecture ingeniously integrates the advantages of edge computing, cloud computing, and mobile computing technologies, achieving efficient processing and precise analysis of industrial big data through advanced deep learning algorithms. The main innovative contributions of this research include: (1) Optimized deployment of improved YOLOv5-based algorithm systems, including YOLOv5-efficient net and YOLOv5-mobile net series lightweight models, which significantly enhance defect recognition capabilities in complex industrial environments; (2) Meticulously designed microservice architecture and distributed message queue system, realizing high availability and dynamic scalability; (3) Pioneering construction of a multi-dimensional intelligent result distribution ecosystem, encompassing cross-platform services (WeChat, email), real-time push notifications (iMessage), and mobile application alerts (Android applications), ensuring seamless and timely delivery of detection results to different end users. In rigorous evaluations on a large-scale self-built steel defect dataset, the system demonstrates good performance across key metrics including detection accuracy, recall rate, and processing speed. Practical deployment has validated the system’s high reliability and robust adaptability in industrial production environments. This research not only provides a feasible technical solution for the intelligent transformation of the steel industry but also establishes a forward-looking technical paradigm for intelligent manufacturing in the context of Industry 4.0, laying a solid foundation for achieving unmanned, high-quality modern industrial production.

This research, positioned at the forefront of industrial intelligence, proposes and implements an innovative intelligent ecosystem for steel surface defect management and prediction. Based on microservice architecture concepts, the system ingeniously integrates advanced frameworks such as Vue.js, Spring Boot, and Django, constructing a full-chain intelligent solution from deep learning detection to machine learning prediction. The study breakthrough proposes a "front-end and back-end collaborative optimization" mechanism, achieving seamless integration of deep learning algorithms and machine learning models through innovative data flow strategies. At the algorithmic level, the system creatively designs an adaptive multi-model ensemble learning framework, integrating algorithms such as LSTM, BP neural networks, Adaboost, and SVM, achieving precise capture of complex nonlinear defect patterns. Meanwhile, by improving the YOLOv5 algorithm, the system has made a qualitative leap in the accuracy and efficiency of steel surface defect detection. At the system architecture level, modular and decoupled design significantly enhances the system's maintainability and scalability, laying a solid foundation for future technological iterations and functional expansions. Large-scale industrial experiments demonstrate that the system has achieved breakthrough progress in key indicators such as detection accuracy, management efficiency, and prediction accuracy, marking a new stage in the intelligent manufacturing of the steel industry. This research not only provides a paradigmatic solution for the digital transformation of the steel industry but also points the way for the intelligent and refined development of manufacturing, holding significant theoretical value and practical significance for promoting the implementation of the Industry 4.0 strategy.

In the context of automated testing in the automotive industry, the accurate and effective detection of target icons on vehicle-machine interfaces is crucial for ensuring the efficiency and precision of automated testing processes. This paper addressed the challenge of effectively detecting vehicle-machine interface icons by proposing a three-stage verification detection algorithm based on machine vision. This paper employed an enhanced template matching technique to achieve precise positioning and initial verification of car and machine icons. Subsequently, a two-stage verification process was implemented, which focuses on detecting local changes using pixel blocks. The three-stage verification was accomplished through a three-channel separation detection of characteristic pixel points based on RGB values. Additionally, this paper utilized the collaborative capabilities of LabVIEW and OpenCV to overcome integration challenges. This joint programming approach bridged the gap between LabVIEW and OpenCV. Experimental results unequivocally underscored the paramount significance of the three-stage verification detection algorithm, rooted in the realm of machine vision. The algorithm’s novel three-stage approach, intricately designed to combat background interference, not only elevated the precision of icon detection but also introduced an ingenious mechanism for classifying detection results. In this regard, it successfully overcome the limitations that have historically hindered traditional machine vision algorithms, ultimately demonstrating its paramount importance in the field of computer vision.

Object 6D pose estimation is a critical task in image processing and various other fields. It involves determining both the three-degree-of-freedom translation and the three-degree-of-freedom rotation information of the target object within the camera coordinate system. I-HS4PCS algorithm is introduced to address the issue of extended execution times in the current pose estimation algorithm. Initially, adapted Harris3D is employed to identify and filter feature points that effectively encapsulate point cloud feature information from the original point cloud dataset. These feature points are subsequently utilized as input data for Super4PCS coarse pose estimation step. Additionally, ikd-Tree data structure and point-wise deletion strategy are leveraged to enhance the ICP fine pose estimation procedure, significantly accelerating the search for the closest points. By maintaining the accuracy of pose estimation while simultaneously boosting algorithmic efficiency, I-HS4PCS have achieved notable improvements. Our experimental validation, conducted using the Stanford University dataset and ShapeNet dataset, demonstrated significant advancements in comparison to other algorithms that employ Super4PCS for coarse pose estimation. Specifically, when compared to ICP, AA-ICP, and Sparse ICP, I-HS4PCS showcased a remarkable improvement in execution time by 80%, 66.8% and 88.3%.

Machine vision is more and more widely used in the field of intelligent transportation. In view of the deficiency of traditional gate sensor recognition system in detecting gate skipping events, an evolutionary neural network based on machine vision is proposed to detect gate skipping behavior(skip the gate to avoid buying tickets). This method extracts four eigenvalues based on the changes of picture pixel values and template matching rate at the gate when the gate skipping event occurs, and then classifies the eigenvalues based on the improved evolutionary neural network by using the global search ability of mind evolutionary algorithm and the strong generalization of BP neural network to determine whether the gate skipping event occurs. The experimental results show that the improved evolutionary neural network algorithm can effectively detect the gate skipping event. It has the advantages of good accuracy, no additional sensors and low cost, and has a good engineering application prospect.