Research On Insulator Detection Technology Based On SSD

Insulator is the most widely used and large number of important electrical components in power system.Its own technological defects and long-term complex external environment make it easy to flashover or drop strings.Once it occurs,it will lead to large-scale power outages.Good target detection method can realize the automatic detection of insulators,and provide a basis for the later stage of insulator condition detection and fault detection.Insulators can be detected quickly and accurately by means of computer vision,which can greatly improve the efficiency of inspection,and is of great significance to the effective and safe operation of power grid.Based on the problem of insulator detection in transmission line inspection images,this paper completes the following work for the existing depth model insulator detection methods:Because there is no data set of insulators in the field of deep learning vision at present,aiming at the problem that existing models can not recognize insulator targets,an image detection database of transmission line insulators is constructed,and the SSD of the depth model is fine-tuned by using this database.The fine-tuned SSD model improves the detection effect of insulators in inspection images,realizes zero breakthrough in identifying insulators,and has good detection results for some simple background insulator images.Aiming at the problem that the detection effect of transmission line insulators is not good under some complex background and low resolution conditions after finetuning SSD model,an adaptive default frame method based on the target characteristics of SSD insulators is proposed.The method ignores the specific proportion of insulators in real life.By centralizing the length-width ratio of insulators in the statistical data,the adaptive change of the length-width ratio in the default frame about the target characteristics of insulator image is realized,which makes the insulators with different width-height ratio and different scales have better detection results.The mAP value of SSD model under VGG-16 Net is further increased to 0.875.The accuracy and performance can basically meet the requirements of power line image automatic realtime processing.In view of the good detection effect of multi-scale model for small targets in images,this paper studies the multi-scale layer depth based on the SSD model,synthesizes the measurement method of layer entropy,and uses the standard of average entropy of candidate regions to combine the appropriate shallow features and deep features of the target object,which significantly improves the accuracy of small target insulators.