基于双通道注意力机制的ResNet果实外观品质分类

ResNet fruit appearance quality classification based on dual channel attention m echanism

摘要:

为了实现对采摘后的果实进行快速、精确的外观 品质分类,并配合分拣生产线完成 果实大规模集中分拣,该研究提出了一种基于改进ResNet 的果实分类方法。首先,将 深度残差神经网络(deep residual neural network,ResNe t) 网络中的残差模块与双通道SE模块(dual channel squeeze-and-exc itation block,DC-SE Block)结合,增强有效的通道特 征并抑制低效或无效的通道特征,提高特征图的表达能力,从而提升识别精度；其次,在原 始ResNet 模型中加入Inception 模块,将果实不同尺度的特征进行融合,增强对较小缺陷 的识别能力；最后,对收集到的4类不同外观品质的果实图像进行数据增强并利用迁移学习 的方法对模型进行初始化。以苹果为例进行的试验结果表明:经过数据集训练之后的改进模 型,在测试集下的准确率达到99.7%,高于原模型的98. 5%；精确率达到99.7%,高于原模 型的98.3%；召回率达到99.7%,高于原模型的 98.7%； 在图形处理器(graphic processing unit,GPU )下的平均检测速度达到32.3 帧/s,略低于原模型的35.7帧/s。与GoogleNet、MobileNet等几种 目前先进的分类方法进 行比较并对不同改进模型进行对比试验 的结果表明,该方法具有良好的分类性能,对解决果 实外观品质的精准分级问题具有重要参考价值。

Abstract:

In order to realize the fast and accurate appearance quality classifica tion of picked fruits,and cooperate with the sorting production line to complete the large-sc ale centralized sorting of fruits,a fruit classification method based on improved ResNet is pro posed in this study. Firstly,the residual module in ResNet network is combined with the dual channel squeeze-and-excitation block (DC-SE Block) to enhance the effective channel features,sup press the inefficient or invalid channel features,and improve the expression ability of t he feature map,so as to improve the recognition accuracy.Secondly,the Inception module is added to the original ResNet model to fuse the characteristics of different scales of fruit,so as to enhance the recognition ability of small defects.Finally,four kinds of fruit im ages with different appearance quality are enhanced,and the model is initialized by transfer learning method .Taking apple as an example,the experimental results show that the accuracy of the improved m odel trained by the data set is 99.7%,which is higher than 98.5% of the original model;The pre cision rate is 99.7%,which is higher than 98.3% of the original model;The recall rate reaches 99.7%,which is higher than 98.7% of the original model; The average detection speed under graphic processing unit (GPU) is 32.3 f rame/s,which is slightly lower than 35.7 frame/s of the original model.Compared with several advanced classification methods such as GoogleNet and MobileNet,and compared with differ ent improved models,the results show that the proposed method has good classification performance,an d has important reference value for solving the problem of accurate classification of fruit appeara nce quality.