Deep Learning-Based Multi-Scale Bubble Detection and Feature Analysis
The morphological characteristics of bubbles in gas–liquid two-phase flow serve as an important basis for the study of heat and mass transfer as well as the design of chemical reaction equipment. These characteristics are widely applied in industrial fields such as petroleum, chemical engineering, e...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | , , , |
| Published: |
2025
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The morphological characteristics of bubbles in gas–liquid two-phase flow serve as an important basis for the study of heat and mass transfer as well as the design of chemical reaction equipment. These characteristics are widely applied in industrial fields such as petroleum, chemical engineering, environmental protection, and energy. Due to the large variation in bubble sizes and their dense distribution, it is prone to missed detections and false detections during the detection process. To address this issue, this paper proposes a multiscale bubble detection algorithm, ATS-YOLO, based on YOLOv10. In the backbone network of the model, an attention mechanism that integrates multiscale features and global contextual information is constructed. Additionally, the Attention-based Downsampling (ADown) module is adopted to enhance the model’s capability to capture multiscale features of bubbles. The spatial and channel reconstruction convolution (SCConv) module is integrated in the neck network of the model to reduce feature redundancy in both the channel and spatial dimensions. The efficient complete intersection over union (EfficiCIoU) loss function is adopted to finely adjust the bounding boxes, enhancing the detection capability for multiscale bubbles. To evaluate the effectiveness of the model, experiments are performed on a public bubble data set as well as a self-built data set captured using a high-speed camera, and bubble morphological parameters are extracted. According to the experimental findings, the enhanced algorithm attains an mAP@0.5 of 97% on the self-built bubble data set and 93.8% on the public bubble data set, fully demonstrating that the proposed model has good detection performance and generalization capability. |
|---|