信息工程与人工智能学院知识库

随着社会的进步与发展，人们对蔬菜、水果等生活必需品的质量越来越重视，需求量也逐渐增大，但由于种种诸如销售环节无标准化、从事蔬菜经营企业对蔬菜质量分级的意识较弱等原因，使蔬菜售卖时品质与价格不能对等，不能做到优质优价。当前，高原夏菜已成为甘肃省的支柱性特色产业，但其依旧采用的是传统的人力分拣或者普通机械分拣装置等方式，这些低效的方式不仅会消耗大量的人力、物力与成本，并且产出的蔬菜质量不能得到保证，从而造成消极影响。随着深度学习热度的不断上涨，研究基于深度学习的高原夏菜质量分级方法对甘肃省的高原夏菜销售也具有重要的意义和广泛的应用价值。因此，本文针对采用传统分级方法所出现的问题，提出了利用深度学习研究高原夏菜质量分级的方法，主要内容如下：

（1）构建高原夏菜数据集。针对目前高原夏菜数据集缺少的问题，建立了一套高原夏菜图像采集装置，采集了结球甘蓝、娃娃菜、菜花、西兰花四种蔬菜作为原始数据集，共计2400张。

（2）扩充原始数据集。为保证采集的高原夏菜数据集在模型训练时能有良好的分级效果，本文首先对图像进行粗分割，排除背景干扰；接着使用数据增强的方法对原始图像进行扩充，使其更有利于后续卷积神经网络模型的训练。

（3）提出了针对高原夏菜质量分级的多尺度融合CA-Ghost-EfficientNet模型。首先构建出轻量化模型，将EfficientNet网络模型中的第一层卷积层替换为Ghost层，同时，在网络的最后一层前嵌入CBAM轻量化注意力模块，使网络模型训练速度加快的同时，也能让网络更加关注细微的特征，检测和定位局部有用的信息，使得相似物种的分析更加精确。接着将网络中MBConv结构中的注意力机制SE模块替换为CA模块，使网络能够同时保留特征的长期依赖关系与精准的位置信息。最后改进网络的优化算法，使用RAdam算法并结合多尺度融合算法，使网络能够保留数据集图像的更多特征，同时也避免了网络陷入局部最优的现象。

本文通过将预处理之后的高原夏菜数据集，在改进的EfficientNet模型与经典神经模型上进行训练，并进行消融对比实验，实验结果表明，本文所提方法的准确率与模型参数量都明显优于其他网络与未改进网络，证明了改进网络模型的有效性与可行性。

With the progress and development of society, people pay more and more attention to the quality of vegetables, fruits and other daily necessities, and the demand has gradually increased, but due to various reasons such as the lack of standardization in the sales link and the weak awareness of vegetable quality grading in vegetable trading enterprises, the quality and price of vegetables cannot be equal, and high quality and price cannot be achieved. At present, plateau summer vegetables have become a pillar characteristic industry in Gansu Province, but they still use traditional manual sorting or ordinary mechanical sorting devices, which will not only consume a lot of manpower, material resources and costs, but also the quality of the vegetables produced cannot be guaranteed, resulting in negative impact. With the rising popularity of deep learning, the study of the quality grading method of plateau summer vegetables based on deep learning is also of great significance and wide application value for the sales of plateau summer vegetables in Gansu Province. Therefore, in view of the problems arising from the traditional grading method, this paper proposes a method to study the quality grading of plateau summer vegetables by deep learning, the main contents are as follows:

(1) Build a dataset of plateau summer vegetables. In view of the current lack of plateau summer vegetable dataset, a set of plateau summer vegetable image acquisition device was established, and four vegetables, kale, baby cabbage, cauliflower and broccoli, were collected as the original data set, with a total of 2400 pictures.

(2) Enrich the original dataset. In order to ensure that the collected plateau summer vegetable dataset can have a good grading effect during model training, this paper first coarsely segmented the image to eliminate background interference. Then, the data augmentation method is used to augment the original image to make it more conducive to the subsequent training of convolutional neural network models.

(3) A multi-scale fusion CA-Ghost-EfficientNet model for the quality grading of plateau summer vegetables is proposed. Firstly, a lightweight model is constructed, the first convolutional layer in the EfficientNet network model is replaced by the Ghost layer, and at the same time, the CBAM lightweight attention module is embedded before the last layer of the network, which accelerates the training speed of the network model and allows the network to pay more attention to subtle features, detect and locate locally useful information, and make the analysis of similar species more accurate. Then, the attention mechanism SE module in the MBConv structure in the network is replaced with the CA module, so that the network can retain the long-term dependence of features and accurate location information at the same time. Finally, the optimization algorithm of the network is improved, and the RAdam algorithm is combined with the multi-scale fusion algorithm, so that the network can retain more features of the dataset image, and also avoid the phenomenon that the network falls into local optimum.

In this paper, the improved EfficientNet model and classical neural model are trained on the preprocessed plateau summer vegetable dataset, and the ablation comparison experiment is carried out, and the experimental results show that the accuracy and number of model parameters of the proposed method are significantly better than other networks and unimproved networks, which proves the effectiveness and feasibility of the improved network model.