Lightweight deep learning with Multi-Scale feature fusion for High-Precision and Low-Latency eye tracking

From 2019, countries worldwide have been negatively affected by the corona virus disease 2019 (COVID-19) in all aspects of social life. The high-tech digital industry represented by emerging digital technologies is still vigorous, and correspondingly, the digital economy has become an important force to promote the stable recovery and re-prosperity of the national economy. The digital economy plays a memorable role in preventing and controlling COVID-19, the resumption of work and production, and the creation of new business formats and models. Urban big data (UBD) involves a wide range of dynamic and static data with high dimensions, but there are no mature and clear data classification and grading standards. Currently, it is urgent to strengthen the security protection of high-value datasets. Therefore, a UBD classification and grading method is proposed based on the lightweight (LWT) deep learning (DL) clustering algorithm. It uses a semi-intelligent path based on partial artificial to form data classification (DC) and hierarchical thesaurus, corpus, rule base, and model base. Subsequently, a big data analysis system is built for unstructured and structured data association analysis based on deep learning, spatiotemporal correlation, and big data technology to improve data value and adapt to multiscenario applications. Meanwhile, with the help of data and graphics processing tool Tableau, the present work analyzes the development status and existing problems of digital resources in China. The results show that although China's digital infrastructure is the top in the world, the trading infrastructure is still only 41.65 percentage points. This shows that China's digital economy still has a lot of room for growth in distribution and trading. The analysis of the ownership of data resources indicates that the scores of China's digital economy in accounting, privacy, and security are very low, only 2.4 points, 5.1 points, and 11 points, respectively. This study has solved the problems of distribution and trade in China's digital economy through research and put forward corresponding suggestions for the current development of China's digital economy market. Hence, a preliminary summary and suggestions are made on the development of China's data resources, to promote the open sharing of data, strengthen the management of data quality, activate the data resource market, strengthen data security, and enhance the vitality of the market economy.

This paper aims to explore the seismic mechanical properties of newly developed fabricated aerated lightweight concrete (ALC) wall panels to clarify the interaction mechanism between wall panels and structures. It first introduces the lightweight deep learning object detection algorithm and constructs a network model with faster operation speed based on the convolutional neural network. Secondly, combined with the deep learning object detection algorithm, the quasi-static loading system is adopted to conduct the repeated loading test on two fabricated ALC wall panels. Finally, the hysteresis load-displacement curve of each test is recorded. The experimental results show that the proposed deep learning algorithm greatly improves the operation speed and compresses the model size without reducing the accuracy. The lightweight deep learning algorithm is applied to the study of the slip performance of the wall plate. The pretightening force of the connecting screw characterizes the slip performance between the wall plate and the structural beam, thereby affecting the deformation response of the wall plate when the interstory displacement increases. The hysteresis curve of the ALC wall panel has obvious squeezing effect, indicating that the slip of the connector can unload part of the external load and delay the damage of the wall panel. The skeleton curve suggests that the fabricated windowless ALC wall panel has higher positive and negative initial stiffness and bearing capacity than the fabricated windowed wall panel. However, the degradation analysis of the stiffness curve reveals that the lateral stiffness deviation of the fabricated windowless ALC wall panel is more obvious. It confirms that the proposed connection method based on the lightweight deep learning model can improve the seismic performance of ALC wall panels and provide reference for the structural analysis of embedding fabricated ALC wall panels. This work shows the important practical value for exploring the application effect of embedded ALC wall panels.

Lightweight deep learning is a subfield of artificial intelligence and machine learning that prioritises efficiency and compactness while developing deep learning models. It is ideal for low-powered mobile phones, embedded systems, and internet-of-things devices due to their speed and low latency. To make lightweight deep learning models, pruning and quantization are used to remove unnecessary parameters and reduce model weight accuracy. Transfer learning is used to fine-tune a pre-trained deep learning model on a smaller dataset. This chapter introduces the fundamentals of lightweight deep learning, including various lightweight models and their applications across different industries.

The current work aims to strengthen the research of segmentation, detection, and tracking methods of stem cell image in the fields of regenerative medicine and tissue damage restoration. Firstly, based on the relevant theories of stem cell image segmentation, digital twins (DTs), and lightweight deep learning, a new phase contrast microscope is introduced through the research of optical microscope. Secondly, the results of DTs method and phase contrast imaging principle are compared in stem cell image segmentation and detection. Finally, a lightweight deep learning model is introduced in the segmentation and tracking of stem cell image to observe the gray value and mean value before and after stem cell image movement and stem cell division. The results show that phase contrast microscope can increase the phase contrast and amplitude difference of stem cell image and solve the problem of stem cell image segmentation to a certain extent. The detection results of DTs method are compared with phase contrast imaging principle. It indicates that not only can DTs method make the image contour more accurate and clearer, but also its accuracy, recall, and F1 score are 0.038, 0.024, and 0.043 higher than those of the phase contrast imaging method. The lightweight deep learning model is applied to the segmentation and tracking of stem cell image. It is found that the gray value and mean value of stem cell image before and after movement and stem cell division do not change significantly. Hence, the application of DTs and lightweight deep learning methods in the segmentation, detection, and tracking of stem cell image has great reference significance for the development of biology and medicine.

Feature extraction is an important task in image-based pattern recognition applications due to a large amount of different features existing in the image and its multiple application areas. Due to this necessity, a very considerable effort has been made by researchers in this direction, leading in many cases to excellent classification results. In this paper, the impact of deep learning techniques on the performance of these systems will be evaluated. For reliable assessment, a contactless palmprint-based biometric system has been developed, which is a typical pattern recognition application. In this study, a simple and lightweight deep learning architecture (ICANet) was used for the feature extraction process. The experimental results of ICANet are compared to other lightweight deep learning (PCANet and DCTNet). The results of the comparison prove the effectiveness. The experimental results of ICANet were compared to lightweight deep learning (PCANet and DCTNet) where the comparison results showed the efficiency of ICANet in terms of classification rate.

With the explosive growth of sports video data on the internet platform, how to scientifically manage this information has become a major challenge in the current big data era. In this context, a new lightweight player segmentation algorithm is proposed to realize the automatic analysis of basketball game video. Firstly, semantic events are expressed by extracting group and global motion features. A complete basketball game video is divided into three stages, and a basketball event classification method integrating global group motion patterns and domain knowledge is proposed. Secondly, a player segmentation algorithm based on lightweight deep learning is proposed to detect basketball players, segment the players, and finally extract players' spatial features based on deep learning to realize players' pose estimation. As the experimental results indicate, when a proposed 2-stage classification algorithm is used to classify the videos, the accuracy of identifying layup, the shooting, and other 2-pointers are improved by 21.26% and 6.41%, respectively. And the accuracy of average events sees an improvement of 2.74%. The results imply that the 2-stage classification based on event-occ is effective. After comparing the four methods of classifying players, it is found that there is no significant difference among these four methods about the accuracy of segmenting. Nevertheless, when judged with the time that these methods take separately, FCN-CNN (Fully Convolutional Network-Convolutional Neural Network) based on superpixels has overwhelming advantages. The event analysis method of basketball game video proposed here can realize the automatic analysis of basketball video, which is beneficial to promoting the rapid development of basketball and even sports.

The purpose is to improve the training effect of physical education (PE) based on the teaching concept of ideological and political courses. The research is supported by the lightweight deep learning (DL) model of the Internet of things (IoT). Through intelligent recognition and classification of human action and images, it discusses the PE and training scheme based on the lightweight DL model. In addition, by the optimization of the accelerated compression algorithm and the evaluation of the PE and training effect of the Openpose algorithm, an optimization model of the PE and training effect has been successfully established. The research data results indicate that after 120 iterations of the model, the system recognition accuracy of the convolutional neural network (CNN) algorithm can only be improved to about 75%, while the recognition accuracy of the Openpose algorithm can reach about 85%. Compared with the CNN algorithm under the same number of iterations, the recognition accuracy can be improved by 9.8%. In addition, when the number of nodes in the network layer is 60, the system delay time of the proposed Openpose algorithm is smaller. At this time, the system delay of the algorithm is only 10.8s. Compared with the CNN algorithm under the same conditions, the proposed algorithm can save at least 1.2s in system delay time. The advantage of the algorithm is that it can improve the efficiency of physical training and teaching, and this research has important reference significance for the digital and intelligent development of the teaching mode of PE.

With the recent success of the deep neural networks (DNNs) in the field of artificial intelligence, the urge of deploying DNNs has drawn tremendous attention because it can benefit a wide range of applications on edge or embedded devices. Lightweight deep learning (DL) indicates the procedures of compressing DNN models into more compact ones, which are suitable to be executed on edge devices due to their limited resources and computational capabilities while maintaining comparable performance as the original. Currently, the approaches of model compression include but not limited to network pruning, quantization, knowledge distillation, neural architecture search. In this work, we present a fresh overview to summarize recent development and challenges for model compression.

To reduce the risk of head trauma to workers working in high-risk workplaces such as construction sites, we designed a new automated lightweight end-to-end convolutional neural network to identify whether all people on a construction site are wearing helmets. Firstly, we used GhostNet as the backbone feature extraction network to take advantage of its low running cost and make the model lighter overall while ensuring efficient automatic feature extraction. Secondly, we designed a multi-scale segmentation and feature fusion network (MSFFN) in the feature-processing stage to improve the algorithm’s robustness in detecting objects at different scales. In contrast, the design of the feature fusion network can enrich the diversity of helmet features and improve the accuracy of helmet detection when distance changes, viewpoint changes, and occlusion phenomena occur. Thirdly, we proposed an improved version of the attention mechanism, the lightweight residual convolutional attention network version 2 (LRCA-Netv2). The main idea of the improvement is implemented around the spatial dimension by fusing the combined features along with the horizontal and vertical directions and then weighting them separately. Such an operation allows the establishment of dependencies between the more distant features with improved accuracy compared to the original LRCA-Net. Finally, when tested on the dataset, the proposed lightweight helmet-wearing detection network has a mAP and FPS of 93.5% and 42.

Chapter 7 - Lightweight deep learning

This work aims to improve the feature recognition efficiency of painting images, optimize the style transfer effect of painting images, and save the cost of computer work. First, the theoretical knowledge of painting image recognition and painting style transfer is discussed. Then, lightweight deep learning techniques and their application principles are introduced. Finally, faster convolutional neural network (Faster-CNN) image feature recognition and style transfer models are designed based on a lightweight deep learning model. The model performance is comprehensively evaluated. The research results show that the designed Faster-CNN model has the highest average recognition efficiency of about 28 ms and the lowest of 17.5 ms in terms of feature recognition of painting images. The accuracy of the Faster-CNN model for image feature recognition is about 97% at the highest and 95% at the lowest. Finally, the designed Faster-CNN model can perform style recognition transfer on a variety of painting images. In terms of style recognition transfer efficiency, the highest recognition transfer rate of the designed Faster-CNN model is about 79%, and the lowest is about 77%. This work not only provides an important technical reference for feature recognition and style transfer of painting images but also contributes to the development of lightweight deep learning techniques.

Synthetic aperture radar (SAR) satellites produce large quantities of remote sensing images that are unaffected by weather conditions and, therefore, widely used in marine surveillance. However, because of the hysteresis of satellite-ground communication and the massive quantity of remote sensing images, rapid analysis is not possible and real-time information for emergency situations is restricted. To solve this problem, this paper proposes an on-board ship detection scheme that is based on the traditional constant false alarm rate (CFAR) method and lightweight deep learning. This scheme can be used by the SAR satellite on-board computing platform to achieve near real-time image processing and data transmission. First, we use CFAR to conduct the initial ship detection and then apply the You Only Look Once version 4 (YOLOv4) method to obtain more accurate final results. We built a ground verification system to assess the feasibility of our scheme. With the help of the embedded Graphic Processing Unit (GPU) with high integration, our method achieved 85.9% precision for the experimental data, and the experimental results showed that the processing time was nearly half that required by traditional methods.

The advancement of information technology has changed traditional manufacturing and business methods, resulting in the emergence of a new business mode known as electronic commerce (E-Commerce). Owing to its obvious benefits, E-Commerce has been extensively employed in a short time, creating a group of E-Commerce enterprises. Establishing financial management strategies that are appropriate for E-Commerce enterprises is critical since it not only aids executors in formulating better financial policies but also benefits enterprises’ administration and market competitiveness. Most of the retail stores in the technological environment are taking different dimensions in their performance through this enterprise E-Commerce. In this study, an E-Commerce system is implemented for retail marketing using lightweight deep learning technology. The deep Lagrangian multiplier approach is used to promote the user’s purchase behavior and to determine whether the estimated optimal transaction quantity is achieved. The user can utilize the mobile application with the internetworking facility to place the order for required products. The proposed system showed the highest performance achieving 98.78% accuracy as compared to the existing system with 92.46% accuracy.

Compressive strength is a vital measure for assessing the quality of high-performance concrete (HPC), and real-time monitoring of its early strength development is essential for guiding construction processes and ensuring structural safety. However, studies on early strength monitoring using the micro-impedance method are relatively scarce. This article introduces a novel approach for early strength monitoring, leveraging a multistacked piezoelectric impedance lightweight monitoring system combined with deep learning techniques. Three primary innovations are presented: (1) this study addresses the challenges of inadequate real-time performance, limited portability, and high cost in traditional structural health monitoring systems by proposing a hardware architecture that integrates a multilayer stacked piezoelectric sensor with the AD5933 impedance analyzer. By incorporating lightweight deep learning algorithms, the study develops a portable impedance monitoring system characterized by high sensitivity, high precision, and easy lightweight deployment. (2) The research innovatively introduces a lightweight deep learning model based on a multigrid architecture, comprising a feature extraction module and a metric learning module. The former achieves hierarchical feature capture through feature fusion of multiscale grid data, cross-modal analysis, and the use of multispecification convolutional kernels. The latter employs graph convolutional network layers to adaptively learn node parameters and optimizes feature parameters via similarity metric aggregation. (3) To address the issue of insufficient data during collection, this study utilizes a signal reconstruction and generation method based on an adaptive orthogonal matching pursuit algorithm. This approach enables the reconstruction of impedance data, mitigating the impact of limited data volume and providing a foundational data framework for subsequent deep learning databases. The proposed method shows promising potential for advancing intelligent early strength monitoring in HPC applications.

Human gait recognition has emerged as a branch of biometric identification in the last decade, focusing on individuals based on several characteristics such as movement, time, and clothing. It is also great for video surveillance applications. The main issue with these techniques is the loss of accuracy and time caused by traditional feature extraction and classification. With advances in deep learning for a variety of applications, particularly video surveillance and biometrics, we proposed a lightweight deep learning method for human gait recognition in this work. The proposed method includes sequential steps–pretrained deep models selection of features classification. Two lightweight pretrained models are initially considered and fine-tuned in terms of additional layers and freezing some middle layers. Following that, models were trained using deep transfer learning, and features were engineered on fully connected and average pooling layers. The fusion is performed using discriminant correlation analysis, which is then optimized using an improved moth-flame optimization algorithm. For final classification, the final optimum features are classified using an extreme learning machine (ELM). The experiments were carried out on two publicly available datasets, CASIA B and TUM GAID, and yielded an average accuracy of 91.20 and 98.60%, respectively. When compared to recent state-of-the-art techniques, the proposed method is found to be more accurate.