Real Time Detection Of Surface Defects With Inception-Based Mobilenet-Ssd Detection Network

Effective surface defect detection are of great significance for the production of high quality products Aiming at real-time detection of surface defect, we propose a reusable and high-efficiency Inception-based MobileNet-SSD method for surface defect inspection in industrial environment. First, convolutional layers for feature extraction used in SSD were replaced by depthwise separable convolutions utilized in MobileNet so that the speed of the network can be faster. Then, the layer in the base network as convolutional feature layer is constructed as Inception which can extract more rich features through multiple convolution combinations of different scales. Finally, predictions from multiple feature maps with different resolutions are combined by the network to naturally handle objects of various sizes. Experimental results on a surface defect dataset containing 2750 images of 5 classes we established confirm that our network has competitive accuracy and is much faster. For 300 x 300 input, ours network achieves 96.1% mAP on DAGM 2007 test at 73FPS on a NVIDIA GTX 1080Ti, outperforming a comparable state-of-the-art FCN model.