.png){kind=icon}
Return
Forensic image classification with active learning and generative adversarial network-based data augmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.dsp.2025.105649.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Crime scene investigation (CSI) image classification is crucial for forensic analysis, as it facilitates the interpretation of evidence and supports criminal investigations. Conventional approaches depend heavily on deep convolutional neural networks (CNNs) trained on large, fully labeled datasets, which are costly and labor-intensive to obtain. Addressing this challenge, we introduce an innovative active learning (AL) approach that improves model performance with fewer labeled samples. Many AL techniques rely on static selection methods. These methods limit adaptability and do not perform well in dynamic environments. To overcome this, we propose an AL framework that integrates deep reinforcement learning (DRL) with a scope loss function (SLF) to dynamically select the most informative unlabeled samples. The SLF balances exploiting known data with exploring new data opportunities. The proposed model improves adaptability by using a GAN with selective gradient omission for dynamic data augmentation. Moreover, to optimize hyperparameters efficiently and stably, it employs an enhanced DE algorithm guided by k-means mutation. Experiments on the CIIP-CSID1, CIIP-CSID2, and CIIP-CSID3 datasets demonstrate that the proposed model achieves average F-measures of 93.62%, 88.45%, and 87.14%, respectively. It outperforms existing methods and reduces the need for extensive manual labeling. These results confirm that combining AL, GAN-based augmentation, and optimized hyperparameters can significantly enhance the accuracy and efficiency of CSI image classification, providing a practical solution for real-world forensic applications.
Journal
D
IF:
3
Papers:
792
Citations:
0
Organization
View more organizations
Cited Papers
No cited papers available