Target Detection of Power Distribution Equipment based on YOLOV3 with Improved Mask Region Convolution Neural Network

The continuous growth of distributed network, distribution network equipment diversifying gradually and it is difficult. The incapability of human to identify the model and type of distributed network equipment accurately which leads frequent safety accidents. However, the existing research are complex to identify the target equipment accurately and efficiently. This research proposed an You Only Look Once Version 3 (YOLOV3) with improved mask Region-based Convolution Neural Network (RCNN) for target detection effectively from the power distribution equipment’s. The dataset is taken through infrared camera rather than handheld thermal imaging camera for electrical equipment in outside substation. Totally, 1263 images are gathered through capturing substation scene at various timestep, distances and angles. It is preprocessed by data augmentation which enhances the size of data. The preprocessed data are provided into YOLOV3+Improved mask RCNN for the detection of targets. The YOLOV3 adopts one neural network to predict bounding box and class possibilities straightly from the whole image in one forward pass. It shortens the detection and enhances the efficiency. The mask RCNN provide accurate pixel-level segmentation for every instance in image. The performance of YOLOV3+Improved Mask RCNN is calculated by metrics like Mean Average Precision (mAP), precision, accuracy, recall and f1-score. The YOLOV3+Improved Mask RCNN achieves $\mathbf{9 5. 2 8 \%} \mathrm{mAP}, \mathbf{9 5. 6 1 \%}$ precision, $\mathbf{9 5. 8 3 \%}$ accuracy, $91.56 \%$ recall and $\mathbf{9 2. 6 7 \%}$ f1-score when compared to YOLOV4, and Improved YOLOV3 + Pixel feature statistics.