基于改进Mask RCNN的工程车辆分割算法

针对输电通道下施工车辆与输电线之间距离难以计算、工程车辆检测精度较低等问题,提出一种改进Mask RCNN的工程车辆分割算法.首先将特征提取网络中的卷积替换为动态卷积,使网络训练时可以根据输入图像及时调整卷积核的大小,有效提高模型性能;然后在网络中添加NAM注意力机制,提高网络对工程车辆的关注度;最后修改特征融合网络为ssFPN,防止特征融合时信息丢失,加强语义融合,提高模型检测精度.对比试验结果表明,与改进前基于ResNet50的Mask RCNN算法相比,改进后算法提高了对工程车辆的检测精度,mAP提高了4.1%,后续处理得到的车辆轮廓精确,证明了改进后算法的有效性.

Engineering Vehicle Segmentation Algorithm Based on Improved Mask RCNN

To address challenges such as the difficulty in calculating the distance between construction vehicles and transmission lines in transmission channels and the low detection accuracy of construction vehicles, an improved Mask RCNN segmentation algorithm for construction vehicles is proposed. Firstly, the convolutional layers in the feature extraction network are replaced with dynamic convolution, allowing the network to adjust the size of convolutional kernels in a timely manner during training, effectively improving model performance. Then, the normality-based attention module (NAM) attention mechanism is added to the network to improve the attention of the network to engineering vehicles. Finally, the feature fusion network is modified to ssFPN to prevent information loss during feature fusion, strengthen semantic fusion and improve the model''s detection accuracy. Experiments show that compared with the original Mask RCNN algorithm based on ResNet50, the improved algorithm enhances the detection accuracy of engineering vehicles, with a 4.1% increase in mAP (mean average precision). The subsequent processing results in precise vehicle outlines, demonstrating the effectiveness of the improved algorithm.