基于卷积神经网络多特征融合的工件识别与检测

针对工业自动化场景中工件识别与检测精度不够高、特征提取困难、多工件定位困难等问题,提出一种基于卷积神经网络多特征融合的工件检测算法。工件检测算法是在一种单次目标检测器算法基础上,新增了特征融合结构,将图像深层信息与浅层信息融合而得以改进,由基础网络、自定义网络、特征融合结构和检测网络四部分构成。实验测试表明,对于200个不同工件组成的图像数据集检测的平均精度达99.2%,优于改进前的96.3%,单张图片检测时间为0.026s,基本符合工业自动化场景中的实时性要求。

Workpiece Recognition and Detection Based on Convolutional Neural Network Multi-Feature Fusion

An algorithm based on convolutional neural network multi-feature fusion is proposed, aiming at solving the problem of low recognition and detection accuracy of work pieces, as well as the difficultly of feature extraction and multi-workpiece location. Based on a Single Shot MultiBox Detection algorithm, the algorithm is improved by adding the future fusion structure, and combining the depth and shallow information of an image, which is composed of basic network, user-defined network, future fusion structure, and detection network. The experimental results show that the average accuracy of image dataset for 200 different workpieces is 99.2%, and better than 96.3% before improved, and single image detection time is 0.026 s, which basically meets the real-time requirements in industrial automation scenarios.