A lightweight detector based on attention mechanism for aluminum strip surface defect detection

Surface defects, which often occur during the production of aluminum profiles, can directly affect the quality of aluminum profiles, and should be monitored in real time. This paper proposes an effective, lightweight detection method for aluminum profiles to realize real-time surface defect detection with ensured detection accuracy. Based on the YOLOv5s framework, a lightweight network model is designed by adding the attention mechanism and depth-separable convolution for the detection of aluminum. The lightweight network model improves the limitations of the YOLOv5s framework regarding to its detection accuracy and detection speed. The backbone network GCANet is built based on the Ghost module, in which the Attention mechanism module is embedded in the AC3Ghost module. A compression of the backbone network is achieved, and more channel information is focused on. The model size is further reduced by compressing the Neck network using a deep separable convolution. The experimental results show that, compared to YOLOv5s, the proposed method improves the mAP by 1.76%, reduces the model size by 52.08%, and increases the detection speed by a factor of two. Furthermore, the detection speed can reach 17.4 FPS on Nvidia Jeston Nano’s edge test, which achieves real-time detection. It also provides the possibility of embedding devices for real-time industrial inspection.

Online surface defect segmentation on aluminum strip production line using a lightweight and efficient model

Small object detection on drone-captured images is a recently popular and challenging task. From the drone’s perspective, the object scale varies significantly, and tiny objects lack distinguishable appearance information in complex backgrounds, which leads to undesired effects. To solve the issue mentioned above, we propose U-YOLO based on the original YOLOv5 model. We first extend the multi-scale feature fusion network and add a detection head for tiny objects. Secondly, we integrate the convolutional block attention model (CBAM) in the detection head to focus on the critical region of the feature map. Lastly, an attention feature fusion module based on contextual information is designed to combine local and global contextual details of small objects and attention mechanisms to enhance the multi-scale feature fusion capability. Experiments on dataset VisDrone2021-DET show that U-YOLO not only improves the detection performance on drone-captured scenarios but also has a real-time detection speed. Compared to baseline model (YOLOv5s), the mAP result of U-YOLO is increased from 19.07 to 24.1%, and the detection speed is 39FPS. It provides a good balance between detection accuracy and speed, promoting the progress of small object detection algorithms on UAV platforms.

Driver distraction behavior causes a large number of traffic accidents every year, resulting in economic losses and injuries. Currently, the driver still plays an important role in the driving and control of the vehicle due to the low level of vehicle automation and the immature development of autonomous driving. Therefore, it is vital to research distraction detection for drivers. However, in realistic driving scenarios with uncertain information, they are still some challenges in efficient and accurate driver distraction detection. In this paper, an improved deep learning model based on attention mechanisms and bi-directional feature pyramid networks (BiFPN) is proposed to identify driver distractions. Firstly, an improved data augmentation strategy is introduced to increase the data diversity to enhance the generalization capability of the model. Secondly, the squeeze-and-excitation (SE) attention mechanism layer is used after the C3 module of the original backbone network to enhance the important feature information and suppress the minor feature information. Finally, the BiFPN module is introduced into the neck network to better achieve multi-scale feature fusion without increasing the calculation amount too much. The experimental results show that the method proposed in this paper has an average mean accuracy rate (mAP) of 0.956 on the test set. Compared to the original model the mAP has improved by 13.2%. The detection speed of the model is 71 frames per second, and the memory occupation is 15.9 MB. This method has the advantages of high recognition accuracy, fast detection speed, and small memory occupation of the model, which are important for achieving engineering deployment.

In some special locations, phone calls are strictly prohibited due to the possibility of significant security hazards. To prevent and reduce this dangerous behavior, we implemented real-time detection of handheld phones on resource-limited devices and developed a novel target detection network called YOLO-PAI. Initially, a pruning algorithm is mainly used to optimize CSPDarknet53 to reduce the number of model parameters while maintaining detection accuracy. Next, the SE attention mechanism module is introduced, and the SRBlock structure is designed to improve the network's attention to the input data and enhance the network feature extraction capability. Then, a new low-dimensional feature extraction branch is added to help the network extract more diverse and richer feature information. Finally, the InceptionV3 structure replaces the original 3 × 3 convolution to reduce the number of model parameters. the YOLO-PAI network can be executed in real-time on an NVIDIA Jetson TX2 embedded device and tested under various illumination and obstacle occlusion conditions. Experimental results show that YOLO-PAI reduces network structure parameters and computational costs while maintaining accuracy. In the Phonehand_Imgs dataset, the network model is shrunk by 190MB and the accuracy of handheld call detection reaches 94%, which is 1.44% higher than YOLOv4. FPS can reach 45 fps, which is 21 fps higher than the original model. On the other two datasets, YOLO-PAI outperformed other popular networks in terms of detection accuracy and speed. In addition, running on the NVIDIA Jetson TX2 embedded device, its detection speed is more than 15 fps faster than other popular networks.

For real applications of pedestrian detection, both detection speed and detection accuracy are important. In this paper we propose a detector based on effective comparison features (ECFs) for simultaneously improving detection accuracy and speed. ECFs are defined as the features helping to improve actual performance. Using only these ECFs as feature candidates for the split nodes of decision trees, our detector can achieve accurate results. As an additional benefit, detection speed is improved by earlier rejection of negative samples. Experiments are conducted using well-known benchmark datasets for pedestrian detection. The experimental results of our ECF detector show that our detection speed is 1–2 orders of magnitude faster than the speed of state-of-the-art algorithms, with comparable detection accuracy.

Surface defects that affect the quality of metals are an important factor. Machine vision systems commonly perform surface inspection, and feature extraction of defects is essential. The rapidity and universality of the algorithm are two crucial issues in actual application. A new method of feature extraction based on contourlet transform and kernel locality preserving projections is proposed to extract sufficient and effective features from metal surface images. Image information at certain direction is important to recognition of defects, and contourlet transform is introduced for its flexible direction setting. Images of metal surfaces are decomposed into multiple directional subbands with contourlet transform. Then features of all subbands are extracted and combined into a high-dimensional feature vector, which is reduced to a low-dimensional feature vector by kernel locality preserving projections. The method is tested with a Brodatz database and two surface defect databases from industrial surface-inspection systems of continuous casting slabs and aluminum strips. Experimental results show that the proposed method performs better than the other three methods in accuracy and efficiency. The total classification rates of surface defects of continuous casting slabs and aluminum strips are up to 93.55% and 92.5%, respectively.

This paper proposes a novel PVF-YOLO model to extract the multi-scale traffic sign features more effectively during car driving. Firstly, the original convolution module is replaced with the Omni-Dimensional convolution (ODconv) and the feature information obtained from the shallow feature layer is incorporated into the network. Secondly, this paper proposes a parallel structure block module for capturing multi-scale features. This module uses the Large Kernel Attention (LKA) and Visual Multilayer Perceptron (Visual MLP) to capture the information generated by the network model. It enhances the representation ability of feature maps. Next, in the process of training, the gradient concentration algorithm is used to optimize the initial Stochastic Gradient Descent (SGD). Under the condition of real-time detection, it improves the detection accuracy. Finally, to improve the robustness of the model, this paper conducts extensive experiments. Tsinghua-Tencent 100K (TT100K), Changsha University of Science and Technology CCTSDB (CSUST Chinese Traffic Sign Detection Benchmark) are used as the training data set. It verifies that the PVF-YOLO method proposed in this paper enhances the detection ability of traffic signs of different scales, and the detection speed and accuracy are better than the original model.

Traditional flame and smoke detection mostly rely on temperature and smoke sensor, but the detection of temperature detector and smoke detector has a certain lag. In order to solve this problem of hysteresis and low accuracy, we propose an improved YOLOV3 algorithm combined with attention mechanism for flame and smoke detection. Firstly, a multi-scene large-scale flame and smoke image dataset is built. The localization and classification of the flame and smoke areas in the image are annotated precisely. The suspected areas of the flame and smoke in the image are obtained by color analysis, so that the suspected areas of the flame and smoke objects are concerned. Then combined with the feature extraction ability of deep network, the problem of flame and smoke detection is transformed into multi-classification and coordinate regression. Finally, the detection model of flame and smoke in multi-scene is obtained. Our experiments show the effectiveness of the improved YOLOv3 algorithm combined with attention mechanism in flame and smoke detection. Our proposed method achieves outstanding performance in the dataset of flame and smoke image. The detection speed also meets the need of real-time detection.KeywordsFlame and smoke detectionAttention mechanismsYOLOv3Deep learning

Concrete surface crack detection plays a crucial role in ensuring concrete safety. However, manual crack detection is time-consuming, necessitating the development of an automatic method to streamline the process. Nonetheless, detecting concrete cracks automatically remains challenging due to the heterogeneous strength of cracks and the complex background. To address this issue, we propose a multi-scale residual encoding network for concrete crack segmentation. This network leverages the U-NET basic network structure to merge feature maps from different levels into low-level features, thus enhancing the utilization of predicted feature maps. The primary contribution of this research is the enhancement of the U-NET coding network through the incorporation of a residual structure. This modification improves the coding network’s ability to extract features related to small cracks. Furthermore, an attention mechanism is utilized within the network to enhance the perceptual field information of the crack feature map. The integration of this mechanism enhances the accuracy of crack detection across various scales. Furthermore, we introduce a specially designed loss function tailored to crack datasets to tackle the problem of imbalanced positive and negative samples in concrete crack images caused by data imbalance. This loss function helps improve the prediction accuracy of crack pixels. To demonstrate the superiority and universality of our proposed method, we conducted a comparative evaluation against state-of-the-art edge detection and semantic segmentation methods using a standardized evaluation approach. Experimental results on the SDNET2018 dataset demonstrate the effectiveness of our method, achieving mIOU, F1-score, Precision, and Recall scores of 0.862, 0.941, 0.945, and 0.9394, respectively.

At present, the detection of surface defects of paper-plastic composite bags still mainly depends on manual visual inspection, which is very inefficient and easy to cause false detection. Based on the algorithm which combines edge detection and adaptive region growth, binary images of the surface defect areas is extracted, six types of shape features and seven types of invariant moment features are defined by calculation as the basis to complete the effective extraction of the surface features, then a paper-plastic composite bags surface defect detection platform is built to complete the defect detection experiment, the defect screening experiment was completed by the rapid screening method of its surface defect image based on gradient projection difference. At the same time, based on Genetic Algorithm (GA), Particle Swarm Optimization, and Grey Wolf Optimization, parameter optimization models of Support Vector Machine were established, respectively. Through experimental comparison, some results show that the defect detection and classification accuracy rate reach 96.83% based on the GA, and the detection speed with the screening method is 3.75 times faster than that without the screening method. Therefore, the reliability of the image classification and screening method for bags surface defects proposed in this study is verified.

At present, the ship detection of optical remote sensing images based on deep learning has made great progress. However, due to the different use scenarios and specific tasks, how to select an appropriate algorithm according to the characteristics of the target and the target priority, so that achieve the detection goal while consider the detection accuracy and speed, still requires relevant research. In this paper, ship detection methods for optical remote sensing images are studied based on deep learning. First, to meet the needs of ship detection research, according to the characteristics of target size and type, datasets of medium and large ships and small target ships are made, and model training and testing are conducted based on Faster R-CNN, YOLOv4, and SSD algorithms respectively. The actual detection performance of the three algorithms under different ship sizes is obtained. The results show that for medium and large targets, Faster R-CNN has the highest precision, the next is YOLOv4, and SSD is the lowest. The detection speed is that SSD is the fastest, the next is YOLOv4, Faster R-CNN is the slowest. For small target ship detection, YOLOv4 has the best detection accuracy and SSD has the fastest detection speed. Faster R-CNN is not as accurate and fast as the other two algorithms. In addition, for different type ships, the detection results of different algorithms also have some differences. In practical applications, different methods should be used to achieve detection by comprehensively considering such factors as target size, target priority, detection accuracy and speed requirements.

Driver distraction behavior is prone to induce traffic accidents. Therefore, it is necessary to detect them to caution drivers in time for traffic safety. In driver behavior recognition, the variety of behaviors and the diversity of the driving environment can have a certain effect on detection accuracy, and the information loss is severe in most existing methods. These make it challenging to improve the real-time accuracy of driver distraction behavior. In this paper, we propose an improved YOLOv7 based on the channel expansion and attention mechanism for driver distraction behavior detection, named CEAM-YOLOv7. The global attention mechanism (GAM) module focuses on the key information to improve accuracy. With the insertion of GAM into the Backbone and Head of YOLOv7, the global dimensional interaction features are scaled up to enable the network to extract key features. Furthermore, In the CEAM-YOLOv7 architecture, the convolution computation has been significantly simplified, which is conducive to increasing the detection speed. Combined with the Inversion and contrast limited adaptive histogram equalization (CLAHE) image enhancement algorithm, a channel expansion (CE) algorithm for data augmentation is presented to further optimize the detection effect of infrared (IR) images. On the driver distraction IR dataset of Hunan University of Science and Technology (HNUST) and Hunan University (HNU), the verification results show that CEAM-YOLOv7 achieved 20.26% higher <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mAP</i> compared to the original YOLOv7 model, and the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FPS</i> reaches 156, which illustrate that CEAM-YOLOv7 outperforms state-of-the-art methods in both accuracy and speed.

With the rapid development of autonomous driving technology, traditional object detection algorithms often face challenges such as the trade-off between model parameters and recognition accuracy, insufficient detection accuracy, and slow detection speed when dealing with vehicle and pedestrian detection tasks in complex scenarios. To address these challenges, we propose a novel lightweight real-time driving scene detection network, FEEP-YOLO. Based on the YOLOv8 algorithm, our network not only maintains high detection accuracy but also achieves faster training speed and fewer model parameters. FEEP-YOLO employs a lightweight network, Faster Net, to significantly reduce model parameters, thereby accelerating detection speed and better meeting real-time requirements. To further improve detection accuracy, we introduce an EMA attention module into the network and optimize the bounding box loss function, adopting the EIOU loss function to enhance target features and reduce background interference. Additionally, to detect small objects more accurately, we add feature layers rich in semantic information to the detection end. Experimental results demonstrate that FEEP-YOLO outperforms YOLOv8 in terms of both detection speed and accuracy on the KITTI and SODA datasets, while significantly reducing the number of model parameters and model size. FEEP-YOLO achieves high detection speed while improving detection accuracy and reducing computational resource consumption, meeting the practical application requirements under limited computing power, memory space and power consumption. The key contribution of FEEP-YOLO lies in its ability to balance accuracy speed, and efficiency, making it a robust solution for real-time autonomous driving applications in complex environments.

Aiming to address the problems of arbitrary orientations, large aspect ratios, and dense arrangements in ship detection, an arbitrary-oriented ship detection method based on RetinaNet is proposed. Our proposed method includes a rotated RetinaNet, a refined network, a feature alignment module, and an improved loss function. First, the rotated RetinaNet achieves rotation detection by using a feature pyramid network, rotated anchors, the skew intersection-over-union (IoU), and skew nonmaximum suppression. Then, the refined network and feature alignment module are introduced to achieve better detection accuracy. Finally, to address the boundary discontinuity, the loss function is improved by introducing the IoU constant factor. Considering the problems with the HRSC2016 dataset, we establish a new dataset with more accurate labels and more images and object samples. Through an ablation study, we thoroughly analyze the validity of the proposed rotated RetinaNet, feature alignment module, and improved loss function. The experimental results show that our method is superior to other state-of-the-art methods.