Research On Fine-Grained Image Classification Algorithm Based On Diversity Feature Fusion Network

Fine-grained image classification tasks involve a more detailed subdivision of subclasses within a specific category of images.Due to the small differences between the categories,it is necessary to pay closer attention to the fine details of the images.This paper addresses the limitations of weakly supervised methods in extracting detailed information and insufficient representation of multi-granularity features.We propose a multi-scale structure and design an interactive attention mechanism to enhance the capturing capability of subtle features and the utilization efficiency of multi-granularity features.The main contributions of this work are as follows:(1)A fine-grained image classification algorithm(MCAF)with multi-dimensional complementary attention feature fusion is proposed.Addressing the characteristics of small inter-class differences and large intra-class differences in fine-grained image classification,we introduce a Complementary Attention Feature Module,which effectively extracts discriminative information from local image regions.By erasing salient features,the network is forced to focus on other discriminative regions.Furthermore,to eliminate redundant noise that may interfere with the classification,the Multi-scale Attention Convolution Module utilizes local features with different receptive field sizes to identify the most effective detail areas,thereby enhancing the network’s sensitivity to discriminative features.Next,the Interactive Multi-layer Perceptron Module models the semantic information between features,enhancing feature diversity through layered interactions between key regions and global features.Finally,the diverse weighted feature interactions aggregate the extracted image features with complementary information from features that exhibit strong discriminative similarity,thereby enabling more accurate identification of subtle differences between categories.(2)A Fine-grained Image Classification Algorithm(CAIF)for Coordinate Attention Interactive Feature Fusion is Proposed.Existing image feature extraction methods often ignore subtle features and may overlook the relationships between features during modeling.Therefore,our proposed method focuses on potential regions and uses coordinate attention to co-encode channel and spatial features,enhancing the perception of feature diversity and promoting the learning of diversified representations.Specifically,we first use the coordinate attention module to generate position-sensitive and orientation-aware dependency-related weights and focus on discriminative features of the interested region through feature mapping.In addition,we enhance the network’s perception of small discriminative features by suppressing salient features.Secondly,the diversity feature fusion module models the interaction of multigranularity features captured in the above process,achieving cross-layer feature significance aggregation and obtaining more diverse complementary information.Finally,we combine the loss functions through collaborative optimization to provide accurate feedback for each module in the network.The proposed method enables end-to-end training without the need for bounding boxes and multi-stage training.The proposed method is validated on four benchmark datasets,including CUB-200-2011,Stanford Dogs,Stanford Cars,and FGVC-Aircraft,and achieves state-of-the-art performance on all of them.Furthermore,ablation studies in this paper validate the effectiveness of each module in the model,reinforcing the reliability and effectiveness of the proposed method.