基于机器视觉的条烟在线分类方法

The online classification method of cigarette based on machine vision

针对条烟在线分类过程中条烟外表面类间差异性小、图像低照度现象所造成的分类精度低、耗时长的问题,提出一种基于机器视觉的条烟在线分类方法。该方法通过线扫描式视觉传感器,搭配同轴线光源采集条烟视觉图像,预处理采用改进的CLAHE算法,基于HSI色彩空间中的亮度分量I,使用亮度分量直方图峰值和均方差量化裁剪参数,自适应增强低照度图像亮度。分类使用ConvNeXt为主干网络,提出特征融合模块,将浅层纹理特征与深层特征进行二次融合,增加局部细节信息,提高条烟分类的准确性;随后提出block加速模块,在block中加入SimAM无参注意力机制并使用结构重参数化,以优化模型并增快推理速度。实验表明,算法的鲁棒性强,分类精确度达到99.99%,能够满足条烟在线分类的要求。

In response to the low inter-class variability of object texture in the online cigarette classification process,as well as the issues of low image brightness and low classification accuracy caused by low lighting conditions,a machine vision-based high-speed cigarette detection method is proposed.This method utilizes a line-scanning visual sensor combined with a coaxial light source to capture visual images of cigarettes.A modified CLAHE algorithm is employed for image preprocessing,which utilizes the brightness component I in the HSI color space to adaptively enhance the brightness of low-lit images by quantifying histogram peaks and mean square deviation clipping parameters of the brightness component.The ConvNeXt network is used as the backbone network for classification,and a feature fusion module is introduced to fuse shallow texture features with deep features,thereby enhancing local detail information and improving the accuracy of cigarette classification.Furthermore,a block acceleration module with the inclusion of SimAM non-parametric attention mechanism and the use of structural reparameterization is proposed to optimize the model and accelerate the inference speed.Experimental results demonstrate the strong robustness and high classification accuracy(reaching 99.99%)of the proposed algorithm,thus meeting the requirements of online cigarette classification.