基于混合域注意力YOLOv4的输送带纵向撕裂多维度检测

针对输送带纵向撕裂目标检测维度单一、模型复杂度高等问题，提出一种高效的MobileNetv3及YOLOv4集成网络输送带纵向撕裂多维度实时检测方法.基于YOLOv4目标识别算法，通过将轻量化网络MobileNetv3代替CSPDarknet53作为骨干网络，结合高效通道域ECA模块和空间域注意力机制（STNet）构建混合域注意力网络（ECSNet），改进了MobileNetv3嵌入ECSNet，并且提升了模型对空间和通道的关注度.引入深度可分离卷积块代替网络中3*3卷积，并将YOLOv4的检测头（Prediction Heads）缩减为2种尺度，轻量化模型降低网络复杂度和训练难度，完成ECSMv3＿YOLOv4模型的搭建，使用K-means聚类6个Anchors预测目标框高宽，提高网络对表面撕裂的检测性能.研制带式输送机多维度智能巡检样机，采集制作输送带多维度面的纵向撕裂数据集，开展网络模型的训练、测试、识别和定位实验.结果表明，提出算法在测试集中的平均识别准确率为97.8%，识别速度为37帧/s，模型的计算量和参数量为4.882 G和8.851 M，通过试验不同的网络模型效果和改变光...

Multi-dimensional detection of longitudinal tearing of conveyor belt based on YOLOv4 of hybrid domain attention

An efficient MobileNetv3 and YOLOv4 integrated network multi-dimensional real-time detection method for longitudinal tearing of conveyor belt was proposed to aim at the problem of single dimension and high complexity of model in the detection of the longitudinal tearing target of the conveyor belt. The lightweight network MobileNetv3 based on the object detection algorithm of YOLOv4 was used to replace CSPDarknet53 as the backbone network of YOLOv4. The ECSNet was constructed by combining efficient channel domain ECA model and spatial transformer network (STNet). The ECSNet was embedded in MobileNetv3 to improve the attention of model to space and channels. The deep separable convolution block was introduced to replace the 3*3 convolution in the network and the Prediction Heads of YOLOv4 were reduced to two scales. The network model was lightened, the complexity and training difficulty were reduced and ECSMv3_YOLOv4 model was built. The K-means was used to cluster six Anchors to predict the height and width of the bounding box, which improved the detection performance of the network for surface tearing. The multi-dimensional intelligent inspection prototype of belt conveyor was developed, the longitudinal tear data set of multi-dimensional surface of conveyor belt was collected and made. The training, testing, identification and positioning experiments of network model were carried out. The results show that the average detection accuracy of the proposed algorithm in the test set is 97.8%, the recognition speed is 37 frame/s and the computational quantity and parameter quantity of the model are 4.882 G and 8.851 M respectively. By testing the effects of different network models and changing the light intensity, the method embodies the advantages of high detection accuracy, fast speed, lightweight and the proposed algorithm has stronger adaptability and anti-interference ability.