Few-shot image classification based on class-irrelevant feature decoupling graph neural network

In the field of computer vision, similarity measurement serves as a pivotal metric for assessing the similarity or dissimilarity of image features. Image classification models utilize the results of these measurements to accomplish classification tasks. However, the inherent limitations of these measurement methods can mislead the model's judgment regarding classification objects. Additionally, the deviation of the target areas caused by the coupling phenomenon between irrelevant backgrounds and classification objects has long been neglected. In this paper, we propose a novel Class-irrelevant Feature Decoupling Graph Neural Network (CFDGNN) to address the bias problem in the model's attention resulting from the limitations of similarity measurement methods and class-irrelevant feature coupling. The proposed framework consists of three modules, including Adaptive Class Salience Channel Weighting Module (ACS), Main Object Focus Spatial Attention Module (MFS) and Feature Decoupling Graph Neural Network (FDGNN). ACS prioritizes high-discrimination salient features while suppressing other salient features that are not relevant to the task. MFS filters out irrelevant noise from the spatial dimension and optimizes the object's texture. FDGNN prevents the model's attention from being misled by simple and irrelevant information. The experimental results on miniImageNet, CIFAR-FS and CUB-200-2011 datasets show that the proposed algorithm can have more accurate classification than other algorithms, meanwhile correct the attention area of the vision model. Additionally, the visualization experiment can save the main texture and filter out irrelevant background textures.