基于协调注意力的花生荚果品质分级

目的:解决花生荚果品质分级过程中模型参数占用内存大、识别精度低、识别速度慢的问题。方法:提出一种基于深度学习和图像处理的花生荚果品质分级方法,在SqueezeNet模型的基础上,通过引入的协调注意力模块(Coordinate Attention)将得到的特征图分别编码成一对方向感知和位置敏感的注意图,加强获取特征图中感兴趣区域信息的能力;采用梯度集中(Gradient Centralization)的策略改进优化算法;优化最末的fire层及卷积层的参数。提出优化模型CG-SqueezeNet,并应用于花生荚果品质分级。结果:与经典模型试验对比,CG-SqueezeNet模型在实际花生荚果图像数据库上的检测准确率为97.83%,参数内存仅为2.52 MB。结论:该方法适合部署在移动终端等嵌入式资源受限设备上,有助于实现对花生荚果品质的实时准确识别。

Classification of peanut quality based on coordinated attention

Objective: This study focuses on solving the problems of large memory consumption, low recognition accuracy and slow recognition speed in the classification process of peanut quality. Methods: A method for classification of peanut quality based on deep learning and image processing was proposed. The Coordinate Attention module was firstly introduced to encode the obtained feature graph into a pair of direction-aware and position-sensitive attention graph, which improved the ability to obtain the information of the region of interest of the graph. Then, Gradient Centralization was used to improve the optimizer. By modifying the parameters of the last fire layer and the convolution layer. An improved model, CG-SqueezeNet, was applied to peanut pod quality grading. Results: The classical convolutional network models VGG16, AlexNet, DenseNet121, ResNet50, Squeezenet were improved, and five different base classifier models were trained by transfer learning. By comparing with the classic model, it was found that the CG-SqueezeNet model could better learn the features of the region of interest in the image. The detection accuracy of the actual peanut pod image database was 97.83%, and the parameter memory was only 2.52 MB. Conclusion: The method is suitable for deployment on embedded resource-limited devices such as mobile terminals, which helps to realize real-time and accurate identification of peanut pod quality.