煤炭输送带是煤矿开采过程中的主要运输设备，在工作过程中不可避免地有大块矸石、锚杆、木板等异物混入，易造成由皮带撕裂、落煤口堵塞导致的重大安全事故。针对井下色彩辨识度低、前后景对比度差及煤炭与异物间遮挡重叠导致物体边缘特征丢失等检测难题，设计了一种融合多尺度特征的轻量化煤炭输送带异物检测方法。首先，基于一种具有压缩-激励模块的残差视觉网络(Residual Vision Transformer with Squeeze-and-Excitation Block，RepViTSEBlock)的架构，融合高效多尺度注意力(Efficient Mult-Scale Attention，EMA)，构建出C2f_RVB_EMA轻量化结构，利用跨空间学习策略与全局特征建模能力，在提升检测精度的同时大幅度减小网络复杂度;其次，将感受野注意力卷积(Receptive Field Attention Convolution，RFAConv)与卷积注意力模块(Convolutional Block Attention Module，CBAM)结合得到 RFCBAMConv，并嵌入到双向特征金字塔网络，通过空间和通道两个维度赋予卷积注意力权重，提高模型对煤炭输送带中异物的关注度，减少计算开销;同时，为了能够精确地识别出多个异物相互堆叠情况下目标的轮廓信息，构建出基于解耦头结构的Detect_SEAM目标检测头;最后，使用Focaler-IoU回归损失函数替换Complete-IoU函数，有效提升了回归框的精度。为避免理想条件对试验造成的影响，采用井下输送带工作的真实图像作为试验数据集。试验结果表明，输送带异物检测模型的平均精度mAP@0。5达到88。20％，相较于基准模型提高了4。60百分点，而参数量与计算量仅为2。51 ×10 6和6。60 × 10 9，有利于在矿井等复杂条件下部署，为煤炭的高效开采运输提供安全预警。
Multi-scale feature fusion method for detecting foreign objects on lightweight coal conveyor belts
To address target detection challenges such as high dust concentration,low color recognition,poor contrast between front and rear views,and the loss of object edge features due to occlusion and overlap between coal flow and foreign objects,we have developed a lightweight foreign object detection method for coal mine conveyor belts that integrates multi-scale features.First,we construct the C2f_RVB_EMA lightweight structure based on the Residual Vision Transformer with Squeeze-and-Excitation Block network architecture,incorporating the characteristics of the Efficient Multi-Scale Attention mechanism.This design utilizes a cross-space learning strategy with global feature modeling capabilities,enhancing detection accuracy while significantly reducing network complexity.Second,we combine the Receptive Field Attention Convolution with the Convolutional Block Attention Module to create the RFCBAMConv,which is integrated into the dual-stream feature pyramid network.This approach adaptively adjusts the receptive field size of each convolution kernel based on the importance of different image regions.It enhances the model's focus on foreign objects in coal mine conveyor belts by simultaneously assigning convolution attention weights along both spatial and channel dimensions,thereby reducing computational overhead.Meanwhile,to accurately recognize contour information when multiple foreign objects are stacked on top of each other,we introduce the Selective Enhancement Attention Module(SEAM)based on a decoupled head structure,resulting in the Detect_SEAM target detection head.Finally,the Focaler-IoU regression loss function is employed to replace the Complete-IoU function,allowing the model to focus on regression samples with varying levels of detection difficulty.This effectively enhances the accuracy of the regression bounding boxes.To ensure that the experiments are not influenced by ideal conditions,actual images taken during downhole conveyor belt operations were used as the experimental dataset.The experimental results demonstrate that the mAP@0.5 for the conveyor belt foreign object detection model reaches 88.20％,which is 4.60 percentage point higher than that of the baseline model.Additionally,the number of parameters and the computational volume are only 2.51 × 10 6 and 6.60 × 10 9,respectively.All metrics outperform those of other mainstream algorithms at this stage,making it more suitable for deployment in the complex conditions found in mines and other mining environments.
万方数据
维普
