基于CNN与SVDD融合的螺栓图像检测方法研究

Research on bolt image detection method based on a fusion of CNN and SVDD

在传统紧固件质量检测行业中,工人负责对产品合格与否进行质量评估,然而人工检测效率低下、易疲劳、误检率高成为制约紧固件行业智能化的关键问题,针对此问题,该文提出一种卷积神经网络(CNN)与支持向量数据描述法(SVDD)融合的螺栓异常检测模型。首先,图像获取装置通过在全方位设置多个摄像头捕捉螺栓的全表面图像信息,图像输入卷积神经网络逐层提取螺栓图像特征,获取螺栓的中高层特征;然后,SVDD作为异常检测分类器进行螺栓缺陷的识别,针对在线获取螺栓缺陷样本的不足导致的样本不平衡问题,提出采用卷积自编码器建立预训练过程,将学习到的权重迁移到深度SVDD模型上作为初始权重。实验结果表明,相比于其他螺栓检测算法,所提融合模型在自主构造的螺栓侧面图像集、头部图像集和底部图像螺栓集上均可取得较优的识别结果,而且所提算法的时间和空间复杂度控制在一定范围内,具有较好的应用价值和市场推广前景。

In the traditional fastener quality inspection industry,workers are responsible for quality assessment of whether the product is qualified or not.However,the low efficiency of manual inspection,easy fatigue,and high false detection rate have become critical problems restricting the intelligence of the fastener industry.This paper proposes a bolt anomaly detection model with the fusion of convolutional neural network(CNN)and support vector data description(SVDD)to address this problem.First,the image acquisition device captures the full-surface image information of the bolt by setting multiple cameras in all directions,and the image is input to the convolutional neural network to extract the bolt image features layer by layer to obtain the middle and high-level features of the bolt;then,SVDD is used as an anomaly detection classifier,and for the problem of sample imbalance caused by the shortage of online acquisition of bolt defect samples,a convolutional selfencoder is proposed to establish a pre-training process,and the learned weights are transferred to the deep SVDD model as the initial weights.The experimental results show that compared with other bolt detection algorithms,the proposed fusion model can achieve better recognition results on the self-constructed bolt side image set,head image set,and bottom image bolt set.The space complexity is controlled within a specific range which has good application value and market promotion prospects.