基于CBAM-ResNet50模型的水果图像识别
Fruit Image Recognition Based on CBAM-ResNet50 Model

针对传统水果分类识别精度低，人工成本高等问题，提出一种基于CBAM-ResNet50模型的水果图像识别算法。首先，采用迁移学习技术，将ImageNet数据集上训练好的权重参数迁移到ResNet50网络水果图像分类模型中，保留卷积层和池化层，去掉分类器，作为主干网络模型；其次，在主干网络模型后添加混合注意力机制模块，根据不同的输入特征分配不同权重，提取有效特征，忽略无关信息。然后，用全局平均池化(GlobalAvgPool2D)替换平均池化，将高维数据转化为低维数据，提高计算效率并简化模型训练过程。最后，添加dropout正则化，随机失活权重参数比例，以确保网络对噪声和异常值的鲁棒性，构建Batch Normalization层对输入数据进行归一化，帮助网络更好地学习数据信息的特征分布，进而提高网络模型性能。把收集到的水果图像按照随机取样的方法划分为80%训练集和20%测试集两部分，采用旋转、平移和裁剪等技术扩充水果图像数据集的多样性和变化性，本文提出CBAM-ResNet50网络模型与MobileNet-v3、VGG16、AlexNet、Xception、ResNet50网络模型的识别效果进行对比，试验结果表明，该模型能够有效识别出几种常见的水果图像，相较于初始网络，识别准确率增加了6个百分点，测试准确率高达99%。为了进一步验证模型性能，分析了基于迁移学习下的数据集扩充与未扩充，添加混合注意力机制对网络模型的影响，由此得出，该研究方法在水果分类识别中具有很好的实践意义。
In response to the traditional fruit classification recognition accuracy and high labor cost, a fruit image recognition algorithm based on the CBAM-ResNet50 model is proposed. First of all, use migration learning technology to migrate the trained weight parameters trained on the ImageNet data set to the ResNet50 network fruit image classification model, retain the convolutional layer and pool layer, and remove the classifier as the main network model; second, in the main network of the main network, After the model, add a hybrid attention mechanism module, allocate different weights according to different input features, extract effective features, and ignore irrelevant information. Then, the average pooling is replaced with global average pooling (GLOBALAVGP- OOL2D), high -dimensional data is converted into low-dimensional data, increasing the calculation efficiency and simplifying the model training process. Finally, add Dropout regularization, randomly loss weight parameters ratio to ensure that the network has a robustness of noise and abnormal values, and builds the Batch Normalization layer to naturalize the input data to help the network better learn the characteristic distribution of data information. , By improve the performance of network models. The collected fruit images are divided into two parts: 80%training set and 20%test set in accordance with the method of random sampling. The diversity and variation of fruit image data sets are expanded by technology such as rotation, translation and cutting. The network model is compared with the recognition effects of MobileNet-V3, VGG16, Alexnet, Xception, and ResNet50 network model. The test results show that the model can effectively identify several common fruit images. Compared with the initial network, the accuracy rate of recognition has increased. 6 percentage points, the test accuracy rate is as high as 99%. In order to further verify the performance