The unmanned aerial vehicle (UAV) inspection technology based on the deep learning has been widely used in the detection of insulator self-explosion defects (ISDs) for power systems. Although existing methods have achieved the high average precision (AP) in general scenarios, they perform poorly in small object scenarios captured by UAVs and lack generalization. Moreover, most of these methods rely on model structures with large parameters and high floating-point operations (FLOPs), and can only be deployed on cloud computing servers. In response to the above issues, this paper proposes a lightweight adaptive you only look once (LA-YOLO) approach of small object detection for ISDs. Firstly, the Faster-C2f module based on the partial convolution (PConv) is introduced into the backbone network to reduce the parameters and FLOPs of redundant channels during gradient diversion. Secondly, the bidirectional adaptively feature pyramid network (Bi-AFPN) dedicated to small object detection layer adaptively learns the weights of feature fusion at different scales from both spatial and channel dimensions. This method improves the effectiveness of the feature information for small-scale ISDs in the process of feature fusion at different scales. Finally, the task and structure dual decoupling head (DD-Head) introduces the spatial aware convolution (SAConv) in the regression branch to extract spatial feature information in both horizontal and vertical directions. The redundant convolution calculations in the classification branch are also reduced, significantly. The experimental results demonstrate that the LA-YOLO not only has lower parameters and FLOPs, but also outperforms existing methods (InsuDet, ID-YOLO, FINet, and BiFusion-YOLOv3) in the average accuracy of small object detection scenarios in ISDs. Our approach is easier to deploy on the UAV and has good application prospects.

Although these methods (e.g., lightweight structural design, model pruning (MP), and model quantization) can reduce the deployment difficulty of deep-learning models in the insulator defect detection, they significantly reduce the detection accuracy. In response to the above issues, this paper proposes a dynamic focused knowledge distillation approach (DFKD) to construct a knowledge transfer path from the large model to the lightweight small model. Firstly, the important sample focusing mechanism introduces dual focus weight factors and adaptive sample matching to encourage the student model to focus on high-quality difficult samples, in order to reduce the adverse effects of low-quality simple samples. Secondly, the adversarial training process of temperature dynamic learning mechanism constructs soft labels of appropriate difficulty based on different stages of distillation training. This helps improve the learning and generalization abilities of the student model towards higher-order knowledge. Finally, this paper combines the DFKD with the MP to establish an insulator defect detection model (DFKD-MP-YOLO) suitable for edge devices with different computing resources. Experiments show that the DFKD method proposed in this paper outperforms existing KD methods in insulator defect detection. Moreover, compared with existing methods (see, e.g., BiFusion-YOLOv3, InsuDet, and ID-YOLO), the DFKD-MP-YOLO not only has a lighter structure, but also achieves higher accuracy and faster speed.

The preventive cultural relics protection is one of the most concerned contents in archaeology, which includes environmental monitoring and accurate prediction of cultural relics diseases. In view of the deficiency of the analysis of cultural relics data and the prediction of cultural relics diseases, a prediction model of immovable cultural relics diseases based on relevance vector machine (RVM) is proposed. The key factors affecting the disease of immovable cultural relics are found out by the principal component analysis method, and the dimension reduction of data is realized; then, the RVM model under the framework of Bayesian theory is constructed, and the super parameters are estimated by the maximum edge likelihood method; finally, the prediction accuracy of the model is compared with the traditional diseases prediction methods. The experiment results demonstrate that the proposed RVM-based immovable cultural relics disease prediction approach not only has the advantages of more sparse model but also has better prediction accuracy than the traditional radial basis function neural network-based and support vector machine-based methods.

This work presents a method for remote sensing object detection (RSOD) based on target feature enhancement and bounding box (Bbox) auxiliary regression. Due to the characteristics of dense distribution, easy feature loss, and difficult Bbox regression (ground truth boxes of medium and small objects in remote sensing images usually only contain a few pixel sizes, making they difficult to regress from the global image), the problem of low accuracy of RSOD arises. Especially, it is easy to lose the features of medium and small objects in RSOD. This work proposes a target feature enhancement module, which enhances the feature transmission of medium and small objects by mapping deep and shallow features. Furthermore, considering that the existence of low-quality labeled data in remote sensing datasets will affect the training process of detection methods, this work proposes a wise intersection over union (Wise-IoU) loss. The Wise-IoU loss focuses on important ordinary-quality labels and improves the overall performance of RSOD. To solve the problem of difficult Bbox regression caused by the small size in remote sensing objects, this paper also proposes a coarse-to-fine Bbox regression model. The new model improves the regression speed and accuracy of medium and small object Bbox by using the auxiliary IoU loss. In addition, the validity and versatility of the method were verified on the Min-DOTA dataset, UCAS-AOD dataset, RSOD dataset, and Min-AI-TOD dataset. The results show that compared with other methods (see, e.g., FCOS, YOLOx, YOLOv7, and YOLOv8), our method has better detection performance and meets real-time detection requirements.