Application of Improved DNN Algorithm Based on Feature Fusion in Fine-Grained Image Recognition

Abstract

Fine-grained image recognition is a research highlight in the computer vision. Compared with traditional image classification tasks, it places more emphasis on distinguishing objects with similar appearance features but belonging to different categories. However, there are common problems with current fine-grained image recognition, namely insufficient feature extraction and feature utilization. To address these issues, an improved DNN image fine-grained recognition method based on feature fusion is proposed. This method solves the insufficient feature extraction through cross bi-linear feature extraction in multi stream networks. Multi-stream networks are used to enhance feature extraction, and cross bi-linear attention mechanisms are introduced to better capture key features in images. In addition, to enhance feature utilization, the study adopts the feature fusion method. According to the findings, the Add feature fusion method using weighted parameters improves accuracy by 3.6% compared with the conventional Concat feature fusion method. The algorithm performs well on the ROC curve, with an AUC value of 0.947. It effectively solves the feature extraction and utilization, promotes the development of this field, and provides reliable and accurate solutions for practical applications.

Keywords

• Fine-grained image recognition
• cross bi-linear
• convolutional neural network
• multi-scale feature fusion
• multi-stream network