Bingtao Hu’s research while affiliated with Hefei University of Technology and other places

Crack detection is an important task to ensure structural safety. Traditional manual detection is extremely time-consuming and labor-intensive. However, existing deep learning-based methods also commonly suffer from low inference speed and continuous crack interruption. To solve the above problems, a novel bilateral crack detection network (BiCrack) is proposed for real-time crack detection tasks. Specifically, the network fuses two feature branches to achieve the best trade-off between accuracy and speed. A detail branch with a shallow convolutional layer is first designed. It preserves crack detail to the maximum and generates high-resolution features. Meanwhile, the semantic branch with fast-downsampling strategy is used to obtain enough high-level semantic information. Then, a simple pyramid pooling module (SPPM) is proposed to aggregate multi-scale context information with low computational cost. In addition, to enhance feature representation, an attention-based feature fusion module (FFM) is introduced, which uses space and channel attention to generate weights, and then fuses input fusion features with weights. To demonstrate the effectiveness of the proposed method, it was evaluated on 5 challenging datasets and compared with state-of-the-art crack detection methods. Extensive experiments show that BiCrack achieves the best performance in the crack detection task compared to other methods.

Defect detection in steel surface is crucial for engineering quality control. Traditional methods for detecting surface defects on steel materials have issues such as low detection accuracy, slow speed, low level of intelligence, and insufficient utilization of images. In response to these challenges, this paper proposes an improved YOLOv8 model for efficient and accurate detection of defects on steel surface. Firstly, we introduce a single-channel adversarial input strategy (AIS) to enhance the utilization of single-channel images and improve the network's detection effectiveness. Secondly, we utilize various attention modules to enhance the Neck and detection head of the network, thereby further improving the network's expressive power and detection performance. Finally, experiments were conducted on three open datasets, achieving a mAP (mean average precision) of 77.3% on the NEU-DET dataset, outperforming YOLOv8 at 74.1%, a mAP of 65.5% on the GC10 dataset, outperforming YOLOv8 at 64.0%, and a mAP of 73.8% on the Magnetic-tile-defect-datasets, outperforming YOLOv8 at 71.2%. Additionally, the average detection speed of this model is 93 frames per second, effectively balancing detection accuracy and efficiency.