The calculation method of spatial attention mechanism is shown in Eq. In this research, the imaginary part of a two-dimensional Gabor perform was used for feature extraction, with the expression of the two-dimensional Gabor function proven in Formula (Eq. On this examine, the spectral power function Eφ(x,y) was used as the response of input image to the Gabor filter. On this research, an adaptive foreground extraction algorithm was proposed primarily based on edge detection, which was used to filter out background interference and receive the foreground rapidly. The Gabor-YOLO algorithm in this study is composed of an adaptive foreground extraction module primarily based on the Gabor operator, an improved YOLO network primarily based on attention mechanism, and a reasoning module based mostly on contextual information. The output outcomes of each stage of the algorithm on this research are proven in Figure 7, during which (a) is the original image taken by UAV, (b) is pretreated after image grey scale and Gaussian filter, (c) is the characteristic determine extracted with Gabor operator, (d) is the fused character figure of various Gabor options, (e) is the foreground determine, and (f) is the ultimate outcomes of the model.



































The general framework of the Gabor-YOLO algorithm is proven in Figure 1. The UAV picture was divided into the image input foreground extraction module, by which the image was first preprocessed by grey scale and Gaussian filtering, then improved after performing characteristic extraction with the Gabor operator, low voltage power line and finally the foreground area was obtained in the picture and enter to the next module. Figure 2. Obtaining foreground areas from pictures. Many of the backgrounds have been filtered out within the picture of the foreground region given by the Gabor algorithm. Considering that the channel attention mechanism can`t acquire the image place info nicely, the spatial attention mechanism was launched to concentrate to the spatial area, and the corresponding function photos of every channel have been calculated and screened. The standard is then extracted, the characteristic vector in every candidate area of the SVM binary classifier is offered in keeping with the target category to classify the corresponding number, and the goal place info is obtained via regression to finish the target detection.



































Girshick (2015) proposed Fast RCNN on the idea of RCNN, which straight inputs the picture into the convolution, and after passing by the ROI pooling layer, the generated region of curiosity is sent to the totally linked layer after which classification of objects is done with the help of SoftMax classifier. Within the RCNN algorithm, the extraction of features and the classification decision are carried out in collection, and the SVM classifier is used for classification, which leads to the drawback of a large amount of calculation. The RCNN first scans the enter image with the selective search algorithm to extract candidate containers, then scales all candidate boxes to a hard and fast pixel measurement by normalization, after which inputs them into the convolutional neural community to unify the size of the characteristic vector. Within the process of UAV taking pictures and transmission, there have been a whole lot of noises within the picture, which weakened the small print of the image. In the images of a low-voltage distribution community, there are a large number of background pixels having a terrific impression on the efficiency of the sting detection algorithm, so it`s not appropriate to straight use the standard edge detection methodology to extract power traces.



































The very last thing you need to do is mess with these high voltage energy strains, but how a lot clearance should you retain between those power traces and your house or office? With a purpose to eradicate its influence on edge extraction as a lot as possible, the image was processed by Gaussian filtering. The initial image collected by the UAV was an RGB colour mannequin, and the grey scale processing was performed on the target picture to scale back the amount of information. It is a pioneering work that introduces CNN into the field of target detection (Girshick et al., 2010). It has essential epoch-making significance and enormously improves the effect of target detection. It can be seen that the detection model proposed in this examine has achieved good coaching effect. The mAP values of energy strains and auxiliary targets of the proposed algorithm are proven in Table 2. As seen from Table 2, the typical accuracy mAP of the proposed algorithm for power traces can reach 93.4%, the typical accuracy of auxiliary objects equivalent to insulators is satisfactory, and the overall common accuracy is 86.6%, indicating that the proposed algorithm has the advantage of high accuracy. Within the inference module, K-means clustering was performed on the coordinates of all auxiliary targets, and after the ability distribution channel was obtained, and the IOU calculation was carried out with the ability line, to acquire the final power strains extraction outcomes.