Research On Intelligent Detection Method Of Strip Steel Surface Based On Deep Learning

Strip steel refers to the plate or coil that has undergone cold rolling or hot rolling processes,which is widely used in industrial production.In iron and steel enterprises,the offline stage of strip steel production is crucial,and it is necessary to accurately detect and recognize surface defects of strip steel and surface mark through advanced technology to provide reliable input for subsequent warehousing,stacking,quality inspection,and other processes.Based on deep learning and swarm intelligence optimization algorithms,this paper improves the recognition rate of surface defects and surface mark on steel strips.The following research work has been specifically carried out:In order to solve the problems of multiple types of surface defects,large size differences,and easy to miss inspection of strip steel,this paper builds a surface defect detection model for strip steel.The model improves Center Net,combining Res Net50 with FPN and adaptive spatial feature fusion methods to enhance multiscale prediction capabilities.Then the model introduces an efficient channel attention mechanism SENet to calculate dynamic weights.In order to reduce the network parameters and computational complexity of the model,the standard convolution in the Res Net50 residual module is replaced by depthwise separable convolution.Finally,CBNet combines CenterNet with HRNet to improve network resolution.In order to improve the image quality of strip steel surface defect detection and ensure the efficiency of the model,this paper proposes a threshold image segmentation method based on improved tuncate swarm algorithm,and applies it to the image preprocessing stage.Firstly,the improved tuncate swarm algorithm combines Tent map and quadratic interpolation algorithm in the population initialization stage to ensure population diversity;Then,the local search convergence is accelerated through the Golden Sine algorithm;After that,the algorithm introduces the Levy flight strategy to enhance the global search range under a certain probability;Finally,the algorithm uses an improved Gauss disturbance algorithm strategy to disturb the optimal position,improving and coordinating local development and global search capabilities.This paper applies the algorithm to image multi threshold segmentation method,which improves the peak signal to noise ratio and enhances the image quality of steel strip surface defects.The text detection and recognition model for strip steel surface mark is studied and constructed.The text detection model is based on the differentiable binarization algorithm network DBNet,which performs network training on the threshold value of each pixel to achieve adaptive binarization and ultimately outputs text candidate boxes.This paper improves DBNet,introduces an efficient channel attention mechanism SENet,and performs adaptive spatial feature fusion to enhance the prediction ability of multiscale targets.The recognition model improves the VGG network of convolutional recurrent neural networks,trains convolution and recurrent neural network jointly,and improves the recognition rate of strip steel surface mark.