基于YOLOv5−SEDC模型的煤矸分割识别方法

Coal and gangue segmentation and recognition method based on YOLOv5-SEDC model

  • 摘要: 现有煤矸分割识别技术参数量大、分类速度慢和识别准确度不高;YOLOv5−seg模型在上下采样操作中易造成图像表面的纹理细节和灰度特征信息丢失,降低煤矸识别效率,且在训练过程中过分侧重全局特征,而忽略了对煤矸识别至关重要的局部显著区域和特征。针对上述问题,提出了一种基于YOLOv5−SEDC模型的煤矸分割识别方法。首先接收包含煤矸形状信息的图像,并利用主干网络进行特征提取,生成特征图;其次在YOLOv5−seg模型中集成SENet模块,以保留煤与矸石表面的纹理细节和灰度特征,避免下采样带来的信息丢失;然后采用不同,膨胀率的空洞卷积策略替代传统卷积核,不仅扩大了模型的感受野,还有效减少了模型参数量;最后分割检测头对融合后的特征进行精细处理,实现对煤矸的精确分割和识别。在大柳塔煤矿实际煤矸分选现场搭建煤矸图像采集实验平台,消融实验结果表明,YOLOv5−SEDC模型的煤和矸石识别的精确率较YOLOv5−seg模型平均提高1.3%,参数量减少0.7×106个,检测速度提高了1.4 帧/s。对比实验结果表明:① YOLOv5−SEDC模型的精确率较YOLOv3−tiny,YOLOv5−seg,Mask−RCNN模型分别提高了10.7%,2.7%,1.9%,达到95.8%。② YOLOv5−SEDC模型的召回率较YOLOv3−tiny,YOLOv5−seg,Mask−RCNN模型分别提高了3.0%,2.1%,0.9%,达到89.1%。③ YOLOv5−SEDC模型的平均精度均值较YOLOv3−tiny,YOLOv5−seg,Mask−RCNN模型分别提高了6.4%,6.3%,1.8%,达到95.5%。④ YOLOv5−SEDC模型的F1较YOLOv3−tiny,YOLOv5−seg,Mask−RCNN模型分别提高了5.2%,4.2%,2.1%,达到92.2%。⑤ YOLOv5−SEDC模型的检测速度较YOLOv3−tiny,YOLOv5−seg,Mask−RCNN模型分别降低了1.9,1.4,2.7 帧/s。可视化结果表明,YOLOv5−SEDC模型对煤和矸石的检测准确度较YOLOv5−seg和Mask−RCNN模型更高,说明了YOLOv5−SEDC模型在煤矸分割识别上具有较好性能。

     

    Abstract: The existing coal and gangue segmentation and recognition technology has a large number of parameters, slow classification speed, and low recognition accuracy. The YOLOv5-seg model is prone to losing texture details and grayscale feature information on the image surface during up and down sampling operations, which reduces the efficiency of coal and gangue recognition. The YOLOv5-seg model overly focuses on global features during training, while neglecting the locally significant regions and features that are crucial for coal and gangue recognition. In order to solve the above problems, a coal and gangue segmentation and recognition method based on YOLOv5-SEDC model is proposed. Firstly, the method receives an image containing the shape information of coal and gangue, and uses the backbone network for feature extraction to generate a feature map. The method integrates the SENet module into the YOLOv5-seg model to preserve the texture details and grayscale features of coal and gangue surfaces, avoiding information loss caused by down sampling. The method adopts a dilated convolution strategy with different dilation rates instead of traditional convolution kernels. It not only expands the receptive field of the model, but also effectively reduces the number of model parameters. Finally, the segmentation detection head finely processes the fused features to achieve precise segmentation and recognition of coal and gangue. A coal and gangue image acquisition experimental platform is established at the actual coal and gangue sorting site of Daliuta Coal Mine. The ablation experiment results show that the accuracy of coal and gangue recognition of YOLOv5-SEDC model is improved by an average of 1.3% compared to YOLOv5-seg model. The parameter quantity is reduced by 0.7×106, and the detection speed is increased by 1.4 frames/s. The comparative experimental results show the following points. ① The accuracy of the YOLOv5-SEDC model is improved by 10.7%, 2.7%, 1.9% compared to the YOLOv3-tiny, YOLOv5-seg, and Mask-RCNN models, respectively, reaching 95.8%. ② The recall rate of the YOLOv5-SEDC model has increased by 3.0%, 2.1%, and 0.9% compared to the YOLOv3-tiny, YOLOv5-seg, and Mask-RCNN models, respectively, reaching 89.1%. ③ The mAP of the YOLOv5-SEDC model has increased by 6.4%, 6.3%, and 1.8% compared to the YOLOv3-tiny, YOLOv5-seg, and Mask-RCNN models, respectively, reaching 95.5%. ④ The F1 value of the YOLOv5-SEDC model has increased by 5.2%, 4.2%, 2.1% compared to the YOLOv3-tiny, YOLOv5-seg, and Mask-RCNN models, respectively, reaching 92.2%. ⑤ The detection speed of the YOLOv5-SEDC model is reduced by 1.9, 1.4, and 2.7 frames/s compared to the YOLOv3-tiny, YOLOv5-seg, and Mask-RCNN models, respectively. The visualization results show that the YOLOv5-SEDC model has higher detection accuracy for coal and gangue than the YOLOv5-seg and Mask-RCNN models. It indicates that the YOLOv5-SEDC model has good performance in coal gangue segmentation and recognition.

     

/

返回文章
返回