Surface Defect Detection of Ultrathin Fiberboard Based on Improved YOLOv8x

Due to significant variations in scale and the predominance of which are small-scale surface defects in Ultrathin Fiberboard (UTFB), manual visual detection remains the primary detection method. However, the efficiency and accuracy of manual visual detection are insufficient to meet the demands of modern UTFB production. Therefore, an improved YOLOv8x algorithm for efficient and high-precision detection of surface defects in UTFB was proposed in this study. Firstly, surface defect images of UTFB were collected from an actual production line and augmented to construct a comprehensive dataset. Then, using YOLOv8x as the baseline model, EfficientNet-ViT (EfficientNet-Vision Transformer) was introduced as the backbone network to achieve efficient feature extraction and improve the accuracy of object detection through improved self-attention mechanism and efficient Transformer architecture. Furthermore, the CIB (Compact Inverted Block) structure was utilized to optimize the C2f module, improving computational efficiency through efficient convolutional operations. Lastly, the WIoU loss function was introduced, with its dynamic focusing mechanism and improved gradient allocation strategy contributing to the detection of small-scale and multi-scale defects. Experimental results show that compared to the baseline YOLOv8x model, the improved YOLOv8x-ECW (YOLOv8x-EfficientNet-ViT-C2fCIB-WIoU) model achieved a 4.4% increase inmAP@0.5. The number of model parameters and floating-point operations (GFLOPS) were reduced by 61.1% and 62.7%, respectively, with a detection frame rate of 48.3 frames per second. The proposed YOLOv8x-ECW model can achieve efficient and accurate detection of surface defects in UTFB, providing technical support for online quality inspection of related wood products.