DeepBP: A bilinear model integrating multi-order statistics for fine-grained recognition. (January 2023)

• Record Type: Journal Article
• Authors: Du, Yinan; Rui, Ting; Li, Hongwei; Yang, Chengsong; Wang, Dong
• Is Part Of: Computers & electrical engineering. Volume 105(2023)
• Publication Date: 2023-01
• Subjects: Fine-grained recognition -- Bilinear pooling -- Multi-layer perceptron -- Multi-orders features; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2022.108432
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: The method of implicit feature interaction with multi-layer perceptron is introduced into fine-grained recognition, which enables the modeling of higher-order feature interactions. The factorized bilinear pooling is compared with multi-layer perceptron. Then, a model that can fuse multi-order statistical information is proposed, and the information of each order complements each other, thereby improving recognition performance. By using the CAM visualization method, how higher-order statistical features improve the model recognition effect is analyzed. Abstract: The bilinear model using second-order statistical features is an important weakly supervised method for fine-grained recognition. Based on this, fusing higher-order statistical features to obtain more discriminant features is an effective approach for improving the performance of the model. However, the existing framework is difficult to fuse higher-order features due to a sharp increase in the number of parameters caused by the increase in fusion order. To address the issue, this paper proposes a DeepBP model composed of a deep network module and a bilinear pooling module. The bilinear module explicitly captures low-order statistical features, and the deep network module implicitly learns high-order features. The two modules are integrated to achieve multi-level information integration. To verify the model's ability, experiments are conducted on the CUB, Cars, and Aircrafts datasets, and the accuracy of 85.6%, 91.6%, and 88.6% is achieved, respectively.