Aiming at the demand for detecting small defects on the surface of cell phone screens with highly reflective characteristics in industrial production, a lightweight detection algorithm for small defects on the surface...
详细信息
ISBN:
(数字)9798350354010
ISBN:
(纸本)9798350354027
Aiming at the demand for detecting small defects on the surface of cell phone screens with highly reflective characteristics in industrial production, a lightweight detection algorithm for small defects on the surface of highly reflective media based on improved YOLOv8 (MGND-YOLOv8n) is proposed, which realizes a great reduction of computational volume and model parameters under the premise of guaranteeing the accuracy and effectively improves the detection speed. This algorithm employs the lightweight structure of MobileNetV3 to substitute the backbone structure of YOLOv8, thereby enhancing portability and realizing the lightweight network model. Additionally, GSConv and Slim-Neck structures are incorporated at the neck end of YOLOv8 to mitigate redundancy and duplicate information. To maintain the detection speed and reduce the complexity of the model, a detection head based on group normalization convolution is devised. Finally, the WIoU-v3 loss function is utilized to account for the quality imbalance of defective samples. Compared with CIoU, this effectively enhances the gradient gain distribution, improves the convergence speed and regression accuracy. Experimental results show that, compared with the original YOLOv8 algorithm, the proposed algorithm improves the average accuracy of mobile phone screen defect detection results by 1.7% at mAP@0.5, reduces the model parameters and computational load by 56.8% and 67.6% respectively, and the detection speed reaches 134.8FPS.
The ENCF-A * (Expansion Neighborhood and Cost Function-Based A * ) algorithm tackles challenges in the traditional A * method such as lengthy paths and numerous turns. It achieves this by expanding the search scope,...
The ENCF-A * (Expansion Neighborhood and Cost Function-Based A * ) algorithm tackles challenges in the traditional A * method such as lengthy paths and numerous turns. It achieves this by expanding the search scope, elongating step distances, and minimizing path deflection angles. The algorithm's efficiency is enhanced through the incorporation of obstacle distribution and a cost function that encourages movement toward the target. Additionally, the introduction of a turn cost function and the removal of redundant nodes prevent unnecessary detours, resulting in more realistic motion paths. Experimental outcomes exhibit significant improvements: ENCF-A * reduces path nodes by 82.61%, decreases deflection angles by 71.96%, and shortens path distances by 9.71%. Consequently, this algorithm notably enhances the efficiency of tracing Automated Guided Vehicles (AGVs) in intricate environments.
暂无评论