YOLO-SE-A: An innovative machine vision approach for low concentration detection of slurry

Abstract

Flocculation-sedimentation of coal slurry is vital for the efficient utilization of coal. However, the detection technology for low concentration clarification layers during the settlement process suffers limitations such as low accuracy and large lag times. Thus, the present work was aimed at developing a new machine vision method for the 0–1000 mg/L concentration range detection for slurry. Firstly, the high-quality video images were acquired with camera focus of 20 cm, distance between light and camera of 10 cm, and light intensity of 35 Klux. The process of Framing→Cropping→Bilateral filtering→CLAHE equalization→Weighted average grayscale→Mosaic augmentation was designed to obtain the feature dataset of slurry. Compared before and after processing, the RMS contrast, Laplacian variance, entropy and training-size of dataset increased 30.4 %, 1.2 %, 12.8 % and 25 %, respectively. Secondly, the combination of Squeeze-and-Excitation and Anchor boxes was embedded into the classical YOLOv5 model to develop a new YOLOv5-SE-A algorithm, which showed good convergence with object loss of 0.0016, classification loss of 0.0038 and bounding box loss of 0.0025. Compared to YOLOv5, the novel algorithm improved performance in Precision, mAP50, and Recall with increasing absolute values of 3.17 %, 3.27 %, 5.47 %, respectively. Thirdly, laboratory validation through accuracy analysis and confusion matrices confirmed the exceptional performance of YOLOv5-SE-A, with a 97.3 % average detection accuracy and a maximum 2.7 % misidentification rate. Compared to conventional YOLO architectures, the proposed model achieves superior metrics (97.38 % precision, 95.69 % mAP, 98.57 % recall) while sustaining real-time processing at 31 fps. Finally, the successful industrial application of YOLOv5-SE-A demonstrated its advantage and application potential.