信息工程与人工智能学院知识库

随着生活水平的提高，年轻一代对多样化需求的偏好显著增长，尤其在水果和蔬菜的质量及品种上。尽管甘肃省的高原夏菜产业被认为是省级特色产业，但是受到人工分拣成本限制，分级标准并没有在销售环节推广。本文利用深度学习技术的进步，提出了一种基于图像识别的高原夏菜质量分级方法，考虑生产环境硬件资源的限制，提出的模型尽量降低时间和空间复杂度，为方法落地实施奠定基础。具体研究内容包括：
（1）构建高原夏菜数据集。为应对高原夏菜图像数据集不足的挑战，设计并实施了一套高原夏菜图像收集装置。利用此装置，成功采集到了包括结球甘蓝、娃娃菜、菜花、以及西兰花在内的四种蔬菜图像，共计构建了含2400张图片的初始数据集。
（2）扩充原始数据集。为提高模型训练的准确性和效率，本研究在扩展原始高原夏菜数据集方面进行了细致工作。首先，通过粗略分割技术去除了图像中的背景干扰，确保数据的质量。随后，借助拍摄扩展和数据增强手段扩充了初始数据集，从而为卷积神经网络模型的训练提供了更为丰富和多样化的图像资源，进一步优化了训练过程的效果。
（3）提出了针对高原夏菜质量分级的多尺度融合ECA-DS-ResNet50模型。这种融合模型结合了多种技术，在ResNet50和MobileNetV2的基础上进行了改进和融合。首先，ResNet50作为主干网络，通过引入ECA（Efficient Channel Attention）注意力机制和多尺度特征提取的ASPP模块，进一步增强了对图像特征的抽取能力。其次，针对ResNet50中7x7卷积核和池化层的计算成本较高的特点，将其3x3标准卷积替换成深度可分离卷积，整合MobileNetV2的DW（Depthwise Convolution）和PW（Pointwise Convolution）卷积机制取代ResNet50的常规卷积，有效降低了计算量。同时，在每层卷积后引入池化层，进一步减少参数量，优化网络结构，大幅减少了时间和空间复杂度，有效提升模型的运算速度。
本研究通过对预处理后的高原夏菜数据集在改进后的ResNet模型以及其他经典深度神经网络模型上进行训练，并通过消融对比实验进行评估。实验结果显示，本文提出的方法在准确率和模型识别速度方面均显著优于其他网络以及原始未改进的网络，这充分证明了我们所改进的网络模型在有效性和可行性方面的显著优势。
 

With the improvement of living standards, the younger generation's preference for diversified needs has significantly increased, especially in the quality and variety of fruits and vegetables. Although the plateau summer vegetable industry in Gansu Province is considered a provincial-level characteristic industry, the classification standards have not been promoted in the sales process due to limitations in manual sorting costs. This article utilizes the advancement of deep learning technology to propose a high-altitude summer vegetable quality grading method based on image recognition. Considering the limitations of hardware resources in the production environment, the proposed model minimizes time and space complexity as much as possible, laying the foundation for the implementation of the method. The specific research content includes: 
(1) Build a dataset for plateau summer cuisine. To address the challenge of insufficient data sets for high-altitude summer vegetable images, a set of high-altitude summer vegetable image collection devices has been designed and implemented. Using this device, four vegetable images including cabbage, baby bok choy, cauliflower, and broccoli were successfully collected, and an initial dataset containing 2400 images was constructed.
(2) Expand the original dataset. In order to improve the accuracy and efficiency of model training, this study conducted meticulous work in expanding the original high-altitude summer vegetable dataset. Firstly, the background interference in the image was removed through rough segmentation techniques to ensure the quality of the data. Subsequently, the initial dataset was expanded using shooting expansion and data augmentation techniques, providing richer and more diverse image resources for the training of convolutional neural network models, further optimizing the effectiveness of the training process.
(3) A multi-scale fusion ECA DS ResNet50 model was proposed for the quality grading of summer vegetables on the plateau. This fusion model combines multiple technologies and cleverly improves and integrates ResNet50 and MobileNetV2. Firstly, ResNet50 serves as the backbone network, further enhancing the ability to extract image features by introducing ECA (Efficient Channel Attention) attention mechanism and ASPP module for multi-scale feature extraction. Secondly, in response to the high computational cost of the 7x7 convolution kernel and pooling layer in ResNet50, the 3x3 standard convolution was replaced with depthwise separable convolution, and MobileNetV2's DW (Depth Convolution) and PW (Pointwise Convolution) convolution mechanisms were integrated to replace ResNet50's conventional convolution, effectively reducing computational complexity. At the same time, introducing pooling layers after each convolution layer further reduces the number of parameters, optimizes the network structure, significantly reduces time and space complexity, and effectively improves the computational speed of the model.
This study trained the preprocessed high-altitude summer vegetable dataset on an improved ResNet model and other classic deep neural network models, and evaluated it through ablation comparison experiments. The experimental results show that the method proposed in this article is significantly better than other networks and the original unimproved network in terms of accuracy and model recognition speed, which fully demonstrates the significant advantages of our improved network model in terms of effectiveness and feasibility.