Metal Surface Corrosion Level Detection Based On Deep Learning

The corrosion detection and corrosion level classification of the metal surface of the distribution box can make timely judgments and carry out relevant anti-corrosion treatments.Therefore,this kind of automatic detection can assist the maintenance personnel of the distribution box.The existing metal surface corrosion detection methods are relatively cumbersome,time-consuming,labor-intensive.Furthermore,it is difficult to accurately identify metal surface corrosion levels since these methods require high technical requirements for operators.In recent years,as one of the representatives of deep learning algorithms,the convolutional neural network has achieved great success in the direction of digital image and video processing.The convolutional neural network provides an idea for our thesis to use deep learning to detect the corrosion level of the distribution box.In this thesis,the metal surface corrosion image samples of the distribution box are from Hubei Electric Power Company.We conducted a detailed analysis of the environment where the distribution box is located and obtained accurate information about the corrosion level of the metal surface of the distribution box.The research goal is to achieve high-accuracy detection of the corrosion level of the metal surface of the distribution box based on a convolutional neural network.The main research contents of our thesis are as follows:Preprocessing of corrosion image samples of the metal surface of the distribution box.This thesis establishes a reasonable image label based on the characteristics of the picture itself and performs data enhancement on the original metal surface corrosion image of the distribution box to obtain the final experimental sample.Find a method for detecting corrosion level of metal surface based on deep learning.According to the basic structure of the convolutional neural network,we build a convolutional neural network model named MS1 Net used to detect the corrosion level of the metal surface of the distribution box.We add the SENet feature extraction module to assign the importance of the corrosion features of the metal surface.Finally,we prove the effectiveness of the proposed network model MS1 Net by a comparative experiment between the loss function and the network.In order to extract more abundant corrosion characteristics of the metal surface of the distribution box and solve the problem of uneven sample types.First,we introduce the concept of the dynamic convolutional network to improve the MS1 Net network model.According to the dynamic convolution theory,we propose the network model MS2 Net.We also use SKNet to transfer the attention mechanism in the network from the channel attention mechanism to the convolution kernel attention mechanism.In this way,we emphasize the importance of the convolution kernel and expand the field of the network to extract more abundant corrosion characteristics of the metal surface of the distribution box.To solve the problem of unbalanced sample types,we improve the cross-entropy loss function and use the Focal Loss loss function to optimize the MS2 Net network model.We then prove that the improved MS2 Net network model has a higher detection accuracy in the detection of the corrosion level of the metal surface of the distribution box through the comparison between the loss function and the network.According to the improved metal surface corrosion level detection method,we developed a high-precision and robust distribution box metal surface corrosion level detection system for the maintenance personnel of the distribution box.