Road object detection is an essential and imperative step for driving intelligent vehicles. Generally, road objects, such as vehicles and pedestrians, present the characteristic of multi-scale and uncertain distribution which puts a high demand on the detection algorithm. Therefore, this paper proposes a YOLOv3 (You Only Look Once v3)-based method aimed at enhancing the capability of cross-scale detection and focusing on the valuable area. The proposed method fills an urgent need for multi-scale detection, and its individual components will be useful in road object detection. The K-means-GIoU algorithm is designed to generate a priori boxes whose shapes are close to real boxes. This greatly reduces the complexity of training, paving the way for fast convergence. Then, a detection branch is added to detect small targets, and a feature map cropping module is introduced into the newly added detection branch to remove the areas with high probability of background targets and easy-to-detect targets, and the cropped areas of the feature map are filled with a value of 0. Further, a channel attention module and spatial attention module are added to strengthen the network’s attention to major regions. The experiment results on the KITTI dataset show that the proposed method maintains a fast detection speed and increases the mAP (mean average precision) value by as much as 2.86 % compared with YOLOv3-ultralytics, and especially improves the detection performance for small-scale objects.

With the rapid development of industrial informatization and deep learning technology, modern data-driven fault diagnosis (MIFD) methods based on deep learning have been receiving attention from the industry. However, most of these methods require sufficient training samples to achieve the desired diagnostic effect, and the scarcity of fault samples in the actual industrial environment leads to the limitation of the development of MIFD methods. In addition, data-driven fault diagnosis methods often need to face cross-load or even cross-domain problems across different devices due to changes in equipment operating conditions and production requirements. In this paper, we design a parameter optimization and feature metric-based fault diagnosis method with few samples, called model unknown matching network model, for the problem of sparse fault samples and cross-domain between data sets in real industrial environments. The method combines both a parametric optimization-based meta-learning network, which extracts optimization information to adapt between different domains, and a metric-based metric learning network, which extracts metric information for similarity discriminations. The experimental results show that the method outperforms the current baseline method for the five-shot fault diagnosis problem of bearings under limited data conditions and achieves an accuracy of up to 94.4 % in cross-device diagnosis experiments from bearings to gas regulators, indicating the feasibility of the method. The features are visualized by T-SNE to show the validity of the model.

To guarantee the stability and safety of industrial production, it is necessary to regulate the behavior of employees. However, the high background complexity, low pixel count, occlusion and fuzzy appearance can result in a high leakage rate and poor detection accuracy of small objects. Considering the above problems, this paper proposes the Enhanced feature extraction-You Only Look Once (EFE-YOLO) algorithm to improve the detection of industrial small objects. To enhance the detection of fuzzy and occluded objects, the PixelShuffle and Receptive-Field Attention (PSRFA) upsampling module is designed to preserve and reconstruct more detailed information and extract the receptive-field attention weights. Furthermore, the multi-scale and efficient (MSE) downsampling module is designed to merge global and local semantic features to alleviate the problem of false and missed detection. Subsequently, the Adaptive Feature Adjustment and Fusion (AFAF) module is designed to highlight the important features and suppress background information that is not beneficial for detection. Finally, the EIoU loss function is used to improve the convergence speed and localization accuracy. All experiments are conducted on homemade dataset. The improved YOLOv5 algorithm proposed in this paper improves mAP@0.50 (mean average precision at a threshold of 0.50) by 2.8% compared to the YOLOv5 algorithm. The average precision and recall of small objects show an improvement of 8.1% and 7.5%, respectively. The detection performance is still leading in comparison with other advanced algorithms.