.png){kind=icon}
.png){kind=icon}
Return
Relaxed Block Diagonal Sparse Embedding for Image Recognition
.png){kind=icon}
PRE
AI
DOI：10.1109/LSP.2025.3606787.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
For pattern recognition tasks, effective representation of key features in data is crucial. Complex and redundant information in practical applications can interfere with robust feature extraction and weaken the expression of discriminative features. Most existing methods focus on directly fitting the label space, and have limited expression of detailed reconstruction of the inherent structure of data. To solve these issues, a novel relaxed block diagonal sparse embedding (RBDSE) algorithm is proposed for image recognition. Specifically, we focus on the global and local structural details of data by imposing $l_{2,1}$ norm and $F$ norm constraints on the projection matrix. In addition, prior knowledge is introduced to construct a relaxed block diagonal matrix to intervene in the representation learning process, enhancing intra-class similarity and inter-class difference of the reconstructed data. Simultaneously, the joint classification regression term fully utilizes the label information. The interactive supervised learning between projection and representation further strengthens the discriminability and interpretability of feature representation. Finally, the performance advantages of RBDSE against other state-of-the-art methods are comprehensively verified on public benchmarks.
Keywords:
Image recognition
sparse representation
subspace embedding
feature extraction
relaxed block diagonal
Journal
I
IF:
3.9
Papers:
600
Citations:
0