Tingting Sui

Aiming at the problems of low detection accuracy, high missed detection rate and poor robustness in complex scenarios for substation meter defect detection, an improved model FAFPN-YOLOv13 based on YOLOv13n is proposed. In order to optimize the bounding box regression process, the model uses the Focal-EIoU loss function. By decoupling error components and enhancing the weight of hard samples, it solves the problems of slow convergence and low defect localization accuracy of traditional loss functions. A Flow-Aligned Feature Pyramid Network (FAFPN) is designed to realize the alignment and aggregation of multi-scale features through semantic flow learning, which improves the model's feature extraction ability for small target defects and low-contrast defects. Experimental results on the self-built substation meter defect dataset show that the mAP50 and mAP50-95 of the FAFPN-YOLOv13 model reach 84.675% and 66.78% respectively, which are 3.173 and 4.711 percentage points higher than those of the YOLOv13n baseline model, and the precision reaches 88.488%. Its performance is significantly better than mainstream lightweight target detection models such as YOLOv5n and YOLOv8n. The training performance curve verifies that the model has good convergence and stability, and the visual analysis of detection results further proves that the model shows obvious advantages in suppressing background interference, improving the detection accuracy of small defects and positioning accuracy, which can meet the requirements of real-time and accurate detection for intelligent inspection of substations.

Defect detection is an important part of the manufacturing process of industrial products. The processing of aluminum profiles requires more accurate and robust defect detection methods. However, various types of defects, small size of the defect pixel area and high defect-background similarity issues pose challenges to existing industrial defect detection methods for aluminum profiles surface defect (APSD). To address these issues, with dual-path parallel attention mechanism (DP-AM), multi-feature fusion mechanism (MFFM) and parallel spatial pyramid pooling fast module (PSPPF), a novel defect detection network, named DMPDD-Net, is proposed in this paper. Specifically, the proposed detection network enhances the feature extraction ability of small-size defects by designing a parallel DP-AM module. Meanwhile, a self-learning factor is set up in the feature fusion formula to construct the MFFM module, which enhances the expression ability of defect features for APSD through multi-channel feature fuse mechanism. Additionally, the PSPPF module is proposed to perform spatial pyramid pooling in You-Only-Look-Once-Version-Eight model (YOLOv8) to reduce the loss of key features. We validate the effectiveness and improvement of our proposed DMPDD-Net by conducting ablation experiments and algorithm comparisons on the Tianchi aluminum profile surface defect dataset (TAPSDD). Our proposed model outperforms the baseline network with a 3.5% increase in mean average precision (mAP) @0.5 and a 2.8% increase in mAP@0.5:0.95, and a 3.2% increase in Formula-One-Score (F1) for the TAPSDD dataset. Our research results indicate that our proposed network is a promising alternative to the current defect detection methods for APSD.

Defect detection in aluminum profiles helps to ensure product quality. However, aluminum defects suffer from variable object scales and high defect-background similarity issues. To address these issues, an improved multiscale semantic enhancement YOLOv5 defect detection method, namely CW-YOLOv5, has been proposed. First, to reduce missed defections, a cross layer link network (CLLN) is introduced into YOLOv5 to capture complete and comprehensive features during the feature extraction process. To accurately detect small-size defects, the weighted feature fusion mechanism (WFFM) is proposed to be added to shallow and high-level feature fusion networks. Finally, the soft non-maximum suppression (Soft-NMS) module is introduced into YOLOv5 to tackle the feature candidate box filtering task to reduce false and missed detections. The proposed defect detection network is applied to the public Tianchi aluminum profile defect dataset (TCAPD), and CW-YOLOv5 gain the improvement of mAP@0.5 by 2.3% compared to the baseline network. Moreover, the experimental results demonstrate that CW-YOLOv5 can effectively detect aluminum surface defects.

Owing to the small size of the defect pixel area and poor defect-background contrast issues in industrial images, noise and missed detection can easily occur. Therefore, automated defect detection is both necessary and challenging. To address these issues, with parallel attention mechanism (PAM) and dual-channel spatial pyramid pooling-fast block (DC_SPPF), a novel defect detection network, namely, PDDD-Net, is proposed in this paper. First, to make the detection network emphasize small defect areas better, the PAM block is proposed to be embedded into YOLOv5 to obtain more low-level visual features and improve the detection accuracy of microdefects. Meanwhile, by fusing multi-channel features, the DC_SPPF block is proposed to replace the raw spatial pyramid pooling-fast block to acquire richer discriminative features of the defect areas. Finally, The soft non-maximum suppression (Soft-NMS) module is used to undertake the feature candidate box filtering task in YOLOv5 to reduce missed detection. Two public datasets are adopted for the model evaluation: the Tianchi aluminum profile defects dataset (APDDD) and the power line insulator dataset (CPLID). The experimental results indicate that the proposed PDDD-Net network exhibits remarkable defect detection performance compared with other related detection methods.

In the beverage, food and drug industry, more and more machine vision systems are being used for the defect detection of Polyethylene Terephthalate (PET) bottle caps. In this paper, in order to address the result of cylindrical distortions that influence the subsequent defect detection in the imaging process, a very fast image stitching algorithm is proposed to generate a panorama planar image of the surface of PET bottle caps. Firstly, the three-dimensional model of the bottle cap is established. Secondly, the relative poses among the four cameras and the bottle cap in the three-dimensional space are calculated to obtain the mapping relationship between three-dimensional points on the side surface of the bottle cap and image pixels taken by the camera. Finally, the side images of the bottle cap are unfolded and stitched to generate a planar image. The experimental results demonstrate that the proposed algorithm unfolds the side images of the bottle cap correctly and very fast. The average unfolding and stitching time for 1.6-megapixel color caps image can reach almost 123.6 ms.

The focus of fine-grained image classification tasks is to ignore interference information and grasp local features. This challenge is what the visual attention mechanism excels at. Firstly, we have constructed a two-level attention convolutional network, which characterizes the object-level attention and the pixel-level attention. Then, we combine the two kinds of attention through a second-order response transform algorithm. Furthermore, we propose a clustering-based grouping attention model, which implies the part-level attention. The grouping attention method is to stretch all the semantic features, in a deeper convolution layer of the network, into vectors. These vectors are clustered by a vector dot product, and each category represents a special semantic. The grouping attention algorithm implements the functions of group convolution and feature clustering, which can greatly reduce the network parameters and improve the recognition rate and interpretability of the network. Finally, the low-level visual features and high-level semantic information are merged by a multi-level feature fusion method to accurately classify fine-grained images. We have achieved good results without using pre-training networks and fine-tuning techniques.