Research On Steel Plate Defect Detection Algorithm Based On SSD Convolutional Neural Network

Target detection technology is one of the core technologies in computer vision technology.With the rapid development of computer technology and processing capabilities,target detection technology based on deep learning has become increasingly mature.As an indispensable raw material for product production,steel plate plays a pivotal role in the modern development process.The quality of its quality directly determines the quality of the product.Due to the diversity of the production environment,there are many types of surface defects on the steel plate.How to quickly detect Defects in the steel plate play an important role in reducing the impact on subsequent products and meeting the increasingly demanding market needs.With the widespread application of deep learning in target detection technology,more efficient,intelligent and accurate steel plate surface defect detection methods become possible.It is very meaningful to study steel plate defect detection based on deep learning.This article adopts the characteristics of steel plate defects and SSD algorithms.The research has realized and improved an SSD-based defect detection algorithm for steel plates.The specific work of the thesis is as follows:(1)First,this paper studies the network structure of the classic target detection algorithm based on convolutional neural network,combined with the consideration of network complexity,network depth,model accuracy,etc.,selects the SSD target detection model as the research Objects are analyzed from the perspectives of network structure,model parameters,feature extraction and training.In order to improve the adaptability of the model in the real environment,the real images in the industrial pipeline are selected as the data set,the images are filtered and denoised,and the defective targets in the images are marked with Label Img software as the training set of the network model And test set.(2)In view of the fact that the defect features in the defect image occupies a relatively small proportion in the entire image,the original SSD target detection network uses the entire feature Map to obtain feature weights in the feature extraction stage,which affects the network training speed and accuracy,and introduces an attention mechanism.The two algorithms of channel and spatial attention are superimposed,which increases the sensitivity of the network to target features and improves the detection accuracy of the network.Experiments show that the improved SSD target detection algorithm performs better in the steel plate defect data,and the detection accuracy is 8.7% higher than the original model.(3)There are many small targets in the steel plate defect target,and the SSD target detection model is small in the detection stage.When the weight of the target semantic feature Map is small,the feature fusion technology is introduced to improve the accuracy of the SSD network for small target feature detection.The experimental results show that the SSD algorithm after feature fusion has a better detection effect in steel plate defect detection,especially for small target defects.Finally,the two improvements were added to the SSD network at the same time to form an attention mechanism + feature fusion SSD model.The experimental results show that the accuracy of the improved model network is13.2% higher than that of the original SSD model.Finally,according to the method proposed in this paper,use PyQT writes a target detection simulation system for steel plate defects,and detects single-type defects and mixed-type defects respectively,and the detection results are good.The realization of the simulation system also proves the effectiveness and implementability of the improved algorithm for defect detection based on SSD convolutional steel plate in this paper.