Abstract: We proposed an improved model of metal surface defect detection method. Firstly,  based on the YOLOv3(you only look once v3) object detection model, a multi-scale convolution parallel structure was used to extract and fuse multi-scale features. Secondly, efficient downsampling was used to maintain the feature information and reduce the computation caused by feature dimension raising. Finally, spatial separable convolution was used to  increase the width and depth of the model while keeping the receptive field unchanged, so that an  improved model YOLOv3I (you only look once v3 inception) with  reduced the amount of model parameters and improved  the performance of the model was obtained. The improved model improved the feature  extraction ability for  complex defects and further reduced the requirements for hardware configuration. The experimental results show that the improved model has significantly improved both accuracy  and calculation efficiency, with an  average accuracy  increase of  about 5% on the public dataset, and about 3% on the bearing dataset provided by the enterprise. The amount of model parameters decreases by more than 20%, and the  floating point computation of the model  reduces by 1.6×10⁹ and 1.2×10¹⁰ times on both two datasets respectively.

Key words: defect detection, feature extraction, multi-scale convolution parallel structure, spatial separable convolution, downsampling