Author
Shuai Zhao
Bio: Shuai Zhao is an academic researcher. The author has contributed to research in topics: Object detection. The author has co-authored 2 publications.
Topics: Object detection
Papers
More filters
22 Oct 2021
TL;DR: Wang et al. as mentioned in this paper proposed an automobile rim weld detection algorithm YOLOv4-head2-BiFPN on the basis of YOLOOv4 algorithm to ensure high accuracy and speed of detection, which does not affect the detection speed by strengthening feature fusion and removing redundant detection heads.
Abstract: At present, the production efficiency of automobile rim in the industrial field is affected by the detection process of automobile rim quality after steel forging. The traditional way is to check welding position manually, which can facilitate the air tightness detection after weld is pressurized. However, this can largely affect production efficiency. By introducing computer vision and image processing, the position of the rim weld can be accurately located, which is more accurate and time-saving. In order to ensure high accuracy and speed of detection, we propose an automobile rim weld detection algorithm YOLOv4-head2-BiFPN on the basis of YOLOv4 algorithm. The experimental results show that, for one thing, it does not affect the detection speed by strengthening feature fusion and removing redundant detection heads. For another, the AP75 of the improved YOLOv4-head2-BiFPN algorithm in the automobile rim weld detection task is 7.7% higher than that of the original YOLOv4 algorithm.
22 Oct 2021
TL;DR: YOLOv4-mini as mentioned in this paper combines the lightweight object detection model with small embedded devices and improves the detection accuracy of automobile rim weld, which can reach 100% and 55.15%, respectively.
Abstract: In order to combine the lightweight object detection model with small embedded devices and improve the detection accuracy of automobile rim weld, this paper proposes YOLOv4-mini based on improved YOLOv4-tiny. Firstly, the lightweight network YOLOv4-tiny is adopted as the main architecture. Secondly, the M-SPP structure is added to obtain the main features of the target and utilize the Mosaic method to enhance the dataset. The K-means + + clustering method is utilized to reset the anchor data belonging to the rim weld target. Finally, the parameters of the pre-trained model are fine-tuned to complete the training of the model. Compared with the original algorithm, in the detection task at 50% and 75% threshold of 4240 rim weld datasets, the mAP index can reach 100% and 55.15%, with the detection speed of 173.5FPS, respectively. Therefore, the improved algorithm gains higher accuracy and efficiency for the edge weld detection task.