基于改进多尺度Retinex图像增强和支持向量机的苹果表面缺陷检测

目的 使用改进多尺度Retinex(Multi-Scale Retinex,MSR)图像增强技术和支持向量机(Support Vector Machine,SVM)分类技术实现苹果表面缺陷检测。方法 对苹果实验图像进行MSR图像增强,消除光照不均匀和表面颜色复杂干扰。对图像增强结果图使用自适应gamma矫正提高光晕区域对比度,并使用基于局部灰度的多阈值比较分割消除光晕现象干扰,获得初步缺陷分割结果。在此基础上,提取苹果梗萼与疤痕的颜色特征,并引入SVM对梗萼和疤痕进行分类,并对梗萼进行剔除，仅保留疤痕作为最终检测结果。结果 将本研究的方法部署到嵌入式设备OpenMV4H7Plus中并经实验证明,梗萼识别准确率达到93.8%,疤痕检测准确率达到92.8%。结论 利用改进MSR图像增强技术和SVM分类技术可以在光照不均匀和颜色信息复杂的苹果表面实现疤痕的有效检测。

Apple surface defect detection based on improved multi-scale Retinex image enhancement and support vector machine

ABSTRACT:Objective To realize the apple surface defect detection using the improvement of multi-scale retinex (MSR) image enhancement technology and support vector machine (SVM) classification technology. Methods The MSR image enhancement method was used to eliminate the interference of the light unevenness phenomenon and the complex color of the surface in apple sample images. Adaptive gamma correction was applied to the image enhancement results to improve the contrast of the halo region, and a multi-threshold comparison segmentation method based on the local grayscale was used to eliminate the interference of the halo phenomenon and obtain preliminary defect segmentation results. On this basis, the color features of the apple stem calyx and scar were extracted, and a support vector machine was introduced to classify the stem calyx and scar. The stem calyx was eliminated, and only the scar was retained as the final defect detection result. Results The method described in this paper was deployed in the embedded device OpenMV4H7Plus and experimentally demonstrated that the peduncle identification accuracy reached 93.8% and the scar detection accuracy reached 92.8%. Conclusion The improved MSR image enhancement technique and SVM classification technique could be used to achieve effective detection of scars on apple surfaces with uneven illumination and complex color information.