快速监督学习在显示器配件分类及识别中的应用

Application of Fast Supervised Learning in Classification and Recognition of Display Accessories

针对常见显示器配件难以快速实时分类识别的问题,以显示器配件图像分类识别为核心,构建一种基于监督学习的显示器配件快速视觉识别系统。通过对生产线上实时采集的显示器配件图像进行低通中值滤波消除图像中的噪声、孤立亮点或暗点,使用高斯算子滤波削弱图像像素灰度变化,使图像表面均匀平滑;使用样本集对监督学习分类器进行6次训练;利用监督学习分类器对显示器配件进行分类识别。基于4种分类识别方法的实验对比结果表明:本文方法采用图像的预处理弥补了监督学习分类器因噪声影响而导致分类识别精度下降的不足,在实时性和鲁棒性方面明显优于其他3种分类识别方法,完成分类识别仅需12.9 ms,每一种配件的识别准确率达到96%以上,分类准确率达到100%,该算法满足显示器配件分类识别的工程应用及实时分拣需求。

Aiming at the problem that common display accessories are difficult to quickly classify and recognize in real time,the image classification and recognition of display accessoriesis takenas the core,and a fast visual recognition system for display accessories based on the supervised learningis built.Firstly,low-pass median filtering is performed on the display accessories images collected in real time on the production line to eliminate the noise,isolated bright spots or dark spots in the image,and Gaussian filter is used to weaken the gray changes of image pixels to make the image surface uniform and smooth.Secondly,the sample set to train the supervised learning classifier for 6 timesis used.Finally,the supervised learning classifier to classify and recognize the display accessoriesis used.The experimental comparison results based on the four classification and recognition methods show that the present method uses the image preprocessing to make up for the insufficient classification andrecognition accuracy of the supervised learning classifier due to the influence of noise.It is significantly better than the other three in terms of the real-time and robustness.It takes only 12.9 ms to complete the classification and recognition,and the recognition accuracy of each accessory is over 96%,and the classification accuracy is 100%.The present algorithm meets the engineering application and real-time sorting requirements of display accessory classification and recognition.