基于残差网络和图像处理的干制哈密大枣外部品质检测

针对目前红枣分级装置检测指标单一，难以实现外部品质综合判别的问题，设计了一款基于残差网络结合图像处理的干制哈密大枣外部品质检测系统。首先，通过深度学习图像分类实现裂纹、鸟啄和霉变缺陷检测，为克服当前残差网络计算量大、复杂度高以及信息丢失的问题，提出了一种改进深度残差网络图像分类方法；其次，根据尺寸与纹理数量的等级差异性，提出了一种阈值检测方法，通过提取干制哈密大枣图像面积、周长、拟合圆半径及纹理数量特征，实现尺寸及褶皱检测。试验结果表明缺陷识别模型和尺寸、褶皱检测模型测试准确率分别达到97.25%、93.75%和93.75%。综合缺陷、尺寸和褶皱3种外部品质指标，通过在线采集图像验证系统测试，外部品质综合检测准确率为93.13%，可初步满足干制哈密大枣品质在线检测装备的生产需求。

Detection Method for External Quality of Dried Hami Jujube Based on Residual Network Combined with Image Processing

In view of the current single detection index of the jujube grading device, and it is difficult to realize comprehensive judgement of external quality, thus a dry Hami jujube external quality detection system based on deep learning and image processing was developed. Firstly, crack, bird peck and mildew defects were detected by deep learning image classification. To overcome the problems of large computation, high complexity and information loss of current residual network, an improved image classification method based on deep residual network was proposed. Secondly, according to the grade difference between size and texture quantity, a threshold detection method was proposed, which can realize the detection of size and fold by extracting the features of area, perimeter, fitting circle radius and texture quantity of dried Hami jujube image. The test results showed that the accuracy of models for detecting defect, size and fold were 97.25%, 93.75% and 93.75%, respectively. Combining three external quality indexes, the detection performance of the system was verified by online image acquisition. After testing, the comprehensive accuracy for detecting external quality was 93.13%, which can initially meet the production requirements of online detection equipment for dried Hami jujube quality. The reearch result can provide theoretical basis and technical reference for the development of rapid nondestructive detection system of dried fruit quality.