A Configuration Approach for Convolutional Neural Networks Used for Defect Detection on Surfaces

The manufacturing industries must guarantee that the products delivered to clients do not have defects, such as irregularities on the surface. To that end, complex systems inspect the products on completion of their manufacturing to detect possible defects. But the design and configuration of these systems is cumbersome, specific for each system, and requires a lot of experience. Currently, there is a trend to build all these systems using Convolutional Neural Networks (CNN), due to the theoretical simplicity of this approach: images of the surface of the products are processed by a trained CNN, which detects defects in the images. But the generation of a well-trained CNN is also a complex process, generally not always properly documented in the literature, and strongly dependent on the application domain. To facilitate the use of CNNs, this paper proposes a configuration approach for CNNs whose objective is the detection of defects on the surface of manufactured products. As an example, the approach is used to configure a CNN to detect surface defects on manufactured rails.