CompoNet with SFEL: A convolutional neural network for identifying low-emissivity coating damage

Abstract

The use of low-emissivity coatings as an infrared (IR) stealth method for weaponry is an important guarantee for the survival of weaponry in the battlefield. The damage on low-emissivity coatings leads to a significant drop in the IR stealth performance of weaponry. Therefore, this paper carries out a research on the intelligent damage identification of low-emissivity coatings based on convolutional neural networks (CNNs). A low-emissivity coating damage dataset with the characteristics of large changes in brightness, distortion of images, and large changes in the target scale is constructed. A composite network (CompoNet) based on feature fusion is designed to extract the detailed features of damage and identify damage types. Compared with VGG16, the damage identification accuracy of CompoNet is improved by 5.49% points. The idea of specific feature extraction layers (SFELs) is proposed. Three SFELs are designed to extract the color, texture, and contour features, which are important basis for judging damage types. The SFEL accelerates the convergence of the model and increases the identification accuracy of the model by 2.02% points. The generative adversarial network is used to generate low-emissivity coating damage images, which solves the problem that the number of damage images collected is small due to practical difficulties. The final damage identification accuracy of the CompoNet fused with SFELs (CompoNet with SFELs) constructed in this paper reaches 99.17%, which is much higher than other generic CNNs.