Image recognition of concrete cracks based on the lightweight DeblurGAN-MobileNet model

Image recognition of concrete surface cracks using neural network models has become an effective method for identifying defects in concrete buildings. However, the accuracy of crack recognition is affected by the fuzzy images acquired during the motion of recognition devices mounted on drones and smart vehicles. The high complexity of deep neural network models limits their application in devices for the intelligent identification of concrete cracks. Therefore, in this work, a lightweight concrete crack image recognition network based on a deblurring generative adversarial network and mobile network (DeblurGAN-MobileNet) model has been designed to effectively improve the accuracy and inference rate of concrete crack image recognition in motion blur background. Firstly, in the feature pyramid network (FPN) of DeblurGAN-V2 for motion deblurring, we adopted an "X"-shaped cross network to improve the internal structure of the FPN. This addresses the issue of uneven resolution contribution and significant information loss in the highest and lowest dimensions during cross-scale feature fusion. Secondly, we incorporated dilated convolutions with different dilation rates into the Bottleneck of the MobileNetV3 image classification network and gradually cascaded them through the network. This not only reduces the computational complexity of the network but also enlarges the receptive field without altering the image dimensions, ultimately enhancing recognition accuracy. The experimental results of motion blur restoration and crack image recognition on different datasets show that our method performs excellently in terms of motion blur removal and recognition accuracy with a motion blur background. Using the GOPRO dataset and a self-made concrete image dataset, the peak signal-to-noise ratio (PSNR) reached 23.51 and 21.95, respectively. Using this method, the recognition accuracy (P) for concrete crack images was 0.889, with an average fast inference speed of only 0.47 seconds per image.