基于改进C3D模型的料仓视频分类识别方法
2025,33(2):161-167
摘要：在自动上料控制系统中，针对传统电感式传感器容易受到外界复杂环境干扰，且需要进行繁琐校准工作等问题，提出了一种基于改进C3D模型的料仓视频视觉分类识别方法。基于实验需求，设计了合作标靶和建立了料仓识别视频数据集；将初始C3D模型作为主干网络进行改进，将该模型第3、4、5层卷积层进行精简，使得模型参数量大幅降低，有利于加快推理速度；在轻量化后的C3D模型上融合SE注意力机制，C3D模型从时空两个维度中提取特征，SE注意力机制可以有效在复杂场景视频帧中找出标靶显著区域，在兼顾时序信息的同时能够高效提取特征进而提高识别准确率。实验结果表明，SE-C3D识别模型准确率达到99.61%，与初始C3D模型相比，准确率提高2.48%，与其他典型三维卷积模型对比，各项性能指标也均有明显提升，对未来智能化上料系统的发展具有重要意义。
关键词：上料系统;3D卷积神经网络;视频分类;SE注意力;模型轻量化
Improve the silo video classification recognition algorithm for C3D model
Abstract：In the automatic feeding control system, in order to solve the problems that the traditional inductive sensor is easy to be disturbed by the external complex environment and needs to carry out tedious calibration work, a visual classification and recognition method of silo video based on improved C3D model was proposed. Based on the experimental requirements, a cooperative target was designed and a video dataset for silo identification was established. The initial C3D model is improved as the backbone network, and the convolutional layers of the 3rd, 4th, and 5th layers of the model are simplified, which greatly reduces the number of model parameters and is conducive to speeding up the inference speed. The SE attention mechanism can effectively find the salient area of the target in the video frame of complex scenes, and can efficiently extract features while taking into account the time series information to improve the recognition accuracy. The experimental results show that the accuracy of the SE-C3D recognition model reaches 99.61%, which is 2.48% higher than that of the initial C3D model, and the performance indicators are also significantly improved compared with other typical 3D convolution models, which is of great significance for the development of intelligent feeding system in the future.
Key words：feeding system; 3D convolutional neural network; video classification; SE module;Model refinement
收稿日期：2023-11-17
基金项目：国家自然科学基金项目(61973139)
