轻量化煤矸目标检测方法研究

Research on lightweight coal and gangue target detection method

  • 摘要: 针对目前基于深度学习的煤矸目标检测方法精度低、实时性差、小目标易漏检等问题,采用轻量化网络、自注意力机制、锚框优化方法对SSD模型进行改进,构建Ghost-SSD模型,进而提出一种轻量化煤矸目标检测方法。Ghost-SSD模型以SSD模型为基础框架,采用GhostNet轻量化特征提取网络代替主体网络层VGG16,以提高煤矸目标检测速度;针对浅层特征图中包含较多背景噪声及语义信息不足问题,引入自注意力模块对浅层特征图进行特征增强,提高对前景区域的关注度,并采用扩张卷积增大浅层特征图的感受野,丰富浅层特征图的语义信息;采用K-means算法对锚框进行聚类,优化锚框尺寸设置,进一步提高煤矸目标检测精度。实验结果表明,基于Ghost-SSD模型进行煤矸目标检测时,平均精度均值较SSD模型提高3.6%,检测速度提高75帧/s,且检测精度与速度均优于Faster-RCNN,Yolov3模型,同时对煤矸小目标具有较好的检测效果。

     

    Abstract: In order to solve the problems of low precision, poor real-time performance and easy missing detection of small targets in the current deep learning-based coal and gangue target detection methods, the SSD model is improved by using lightweight network, self-attention mechanism and anchor frame optimization method to construct Ghost-SSD model, and then a lightweight coal and gangue target detection method is proposed.The Ghost-SSD model is based on the SSD model, and the GhostNet lightweight characteristic extraction network is used to replace the main network layer VGG16 so as to improve the detection speed of coal and gangue targets.In order to solve the problem that the shallow characteristic map contains more background noise and insufficient semantic information, the self-attention module is introduced to enhance the characteristics of the shallow characteristic map and increase the focus on the foreground region.Moreover, the dilated convolution is applied to increase the receptive field of the shallow characteristic maps and enrich the semantic information of the shallow characteristic maps.The K-means algorithm is used to cluster the anchor frames, optimize the size of the anchor frame, and further improve the precision of coal and gangue target detection.The experimental results show that when the Ghost-SSD model is applied in coal and gangue target detection, the mean average precision is 3.6% higher than that of the SSD model, the detection speed is increased by 75 frames/s, and the detection precision and speed are better than that of the Faster-RCNN and Yolov3 models.Moreover, the model has a good detection effect on small coal and gangue targets.

     

/

返回文章
返回