Xu Tian

In the context of resource constrained embedded terminals, the current deep learning-based technology for detecting surface defects in strip steel faces challenges with low detection accuracy and efficiency. This study introduces a lightweight convolution neural network (CNN) model for detecting surface defects of strip steel, which is based on deep learning. Through the integration of non-parameter attention mechanism SimAM, inverse residual structure, and depth-wise convolution module, a lightweight attention mobile backbone network is developed to achieve optimal feature extraction. To address the challenge of detecting small surface defects the context augmentation module is introduced to provide more information for small defects detection by using multi-scale features. To improve the efficiency of feature fusion and reduce parameter redundancy, the GVFPN neck network is proposed. The network aims to represent and deal with multi-scale features objectively while minimizing costs. The experimental results indicate that on the NEU-DET dataset, the proposed network attains an mAP of 78.8% and FPS at 97 frame/s. Moreover, the model requires only 2.39 M parameters and 3.1 G FLOPs. Compared to YOLOv5s, the proposed network reduces parameters by 65.9% and FLOPs by 80.7%, while achieving a 1.3% higher mAP and a 30 frame/s increase in detection speed. These results effectively demonstrate the effectiveness and superiority of the proposed method.

This study introduces a novel data augmentation method aimed at addressing the limited availability of sample data for strip steel surface defects, utilizing an improved Auxiliary Classifier Generative Adversarial Network (ACGAN). Firstly, on the generator side, the improved ACGAN integrates an encoder that extracts essential features from images of strip steel surface defects and transforms them into latent vectors. These vectors serve as inputs for the generator, helping it learn the actual data distribution of strip steel defects more accurately. To enhance image resolution and reduce checkerboard artifacts common with transposed convolution, the generator uses a combination of transposed convolution and PixelShuffle for upsampling. Secondly, on the discriminator side, the convolutional network employs an alternating configuration of downsampling and standard convolutional layers, effectively reducing feature dimensions while preserving more local information. A Dropout regularization strategy is applied after each convolutional layer to boost the model's robustness. Finally, the stability of the learning process is reinforced by employing Kaiming initialization for both the generator and discriminator's convolutional and transposed convolution layers. Performance evaluations conducted using the Northeastern University surface defect database confirm the efficacy of the proposed method in generating accurate images of strip steel surface defects in small-sample scenarios. Notably, compared to the original ACGAN and other classical generative adversarial networks, the improved method demonstrated significant improvements, achieving reductions of 52.82% and 38.78% in FID and MMD metrics, respectively.