摘要
为了提高木材的使用效率、避免由于木材缺陷造成生产故障,根据木材缺陷类型对其分类处理是一种有效的手段,但木材缺陷复杂多样且具有诸多相似性使得类别区分成为难点。针对以上问题本文提出了一种基于子区域变尺度高斯拟合模型的缺陷识别方法。首先建立变尺度高斯拟合基本模型,然后将缺陷纹理分成若干子区域,提取各分区的高斯拟合特征并进行融合;将高斯融合特征及圆度和边缘直线度这两个几何特征输入到建立好的BP神经网络模型中进行训练,根据优化训练的网络模型识别缺陷。该方法对自建的SUT-W图库中雪糕棒图像上人工标定的裂缝、矿物线、矿物块和黑节子的准确识别率分别为91.72%、92.77%、92.67%和92.80%,与其他典型纹理检测方法相比,4种缺陷准确识别率最高分别提高9.38%、6.69%、13.55%和10.22%,说明本文方法能够有效地将以上4种缺陷分辨开,具有一定的实际应用价值。
In order to improve the usage efficiency of wood and avoid production failure caused by wood defects, the classification processing according on the type of wood defects is an effective method. However, the wood defects are complicated, diverse and have plenty of similarity, which makes the type classification difficult. Aiming at this problem, a defect identification method based on sub- region zoom Gaussian fitting model （SZGFM） is proposed in this paper. Firstly, the basic model of zoom Gaussian fitting is established. Secondly, the defect texture images are divided into several sub-regions, and the Gaussian fitting feature of each region is extracted, then these features are fused. The fused Gaussian fitting feature together with two geometric features of the roundness and edge linearity are sent to the established BP neural network for training. The defects are identified according to the optimizingly trained network model. The experiments on the manual calibrated defects, such as splits, mineral streaks, mineral blocks and black knots in the self-buih SUT-W image gallery were conducted; and the accurate recognition rates for the four kinds of defects are 91.72%, 92.77%, 92.67% and 92.80%, respectively, and the accurate recognition rates of the four kinds of defects are improved by 6.83%, 6.54%, 4.55% and 3.22%, respectively at most compared with other typical texture feature classification methods. The result shows that the proposed method can distinguish the four kinds of defects effectively and has practical application value.
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2016年第4期879-886,共8页
Chinese Journal of Scientific Instrument
基金
国家自然科学基金(61271365)项目资助
