| In the process of manufacturing,the surface defects of bearing components caused by equipment accuracy deviation and environmental impurities usually have characteristics similar to the background texture,and the size and density of target defects are different,making it difficult to detect industrial metal surface defects.In response to this,this thesis uses the following two ideas to improve model’s performance.First,the basic component unit residual module of YOLOv3 is improved,and the residual module embedded with SA attention is constructed.The YOLOv3-SA model embedded with attention is proposed to correct and enhance the feature extraction ability of the model to a certain extent,and improve the representation capability of target feature.Experimental results on the bearing raceway groove surface defects dataset provided by the bearing manufacturer show the m AP of YOLOv3-SA is increased by about 7% to 95.4%,on the NEU public steel plate surface defects dataset,the m AP of YOLOv3-SA is increased by about 5% to 95.3%.Second,in order to improve the multi-scale feature expression ability of the model for recognition accuracy,the Inception design idea is utilized for proposing YOLOv3 I,the multi-scale convolution parallel structure is used to extract and fuse multi-scale information,and the feature fusion ability of the YOLOv3 I is improved.The ability to use spatially separable convolutions increases the width and depth of the model at the same time,and saves parameters and reduces the amount of model computation through 1×1 convolution.An efficient downsampling method is used to increase feature dimensions and pooling operation at the same time,which reduces the loss of feature information and saves the computation amount caused by increasing dimensions.Compared with YOLOv3,on the bearing dataset,the m AP of YOLOv3 I is increased by about 3% to 94.1%.On the NEU dataset,the average precision of most categories is improved,the m AP of YOLOv3 I is increased by 5% to 84.7%.while the number of model parameters is reduced by about 24%,the computation cost is reduced by 8.7E+09FLOPs.The attention mechanism aims to find and strengthen relevant features from numerous sparse features,and from another perspective,other relatively irrelevant features are weakened.The idea of Inception is to extract multi-scale features of dataset,extract and learn as many sparse features as possible,and learn target features more comprehensively.So attention and Inception are improvement ideas in two different directions. |
