Research on image classification based on fusion of global features and local features

Image classification has always been a classic research topic in the field of image processing. In recent years, the development of deep learning has triggered a new wave of artificial intelligence research, further promoting the development of computer vision related technologies. The leaf image recognition method based on deep learning has become a research hotspot, such as convolutional neural networks that are good at solving image recognition problems, such as LeNet-5, AlexNet, VGGNet, GoogLeNet, ResNet, DenseNet and other neural networks, constantly refreshing their achievements in classification tasks. Most deep learning networks focus on optimizing network structure, while ignoring the complementarity of different receptive fields. Based on this, this article focuses on the complementarity between global and local features of images, and proposes the FGL network model. FGL extracts image features from both large and small receptive fields, and combines their complementarity to propose a hierarchical fusion module, which enhances the extraction of effective image features and improves classification accuracy. FGL achieved better results in CIFAR10, CIFAR100, and SVHN. Reached 96.8%, 84.2%, and 97.9% respectively. FGL has the following advantages: it is an end-to- end network; More feature information can be extracted under the same data conditions.