An off-policy deep reinforcement learning-based active learning for crime scene investigation image classification

Crime scene investigation (CSI) image classification is crucial in forensic analysis, significantly boosting the efficiency of police investigations. Conventional CSI image classification detection approaches depend heavily on convolutional neural networks (CNNs) and pre-labeled, extensive image data, which are time-consuming and costly to assemble. To address this issue, we present an active learning method that boosts model performance using fewer labeled examples. Unlike traditional active learning methods that employ heuristic selection techniques independently of the training process and compromise their effectiveness, our strategy integrates off-policy deep reinforcement learning (DRL) to make strategic data selections. The off-policy approach allows for learning from a broader range of experiences, improving adaptability and accelerating the learning process compared to on-policy methods. Our model utilizes numerous CNNs to pull features from different layers of images, which are subsequently processed by the softmax layer for image categorization. Starting with a minimal labeled dataset, the classifier employs DRL to identify which unlabeled images should be annotated next. These newly labeled images are then added to the training pool, and the classifier is periodically retrained to enhance its accuracy. Additionally, our framework incorporates a generative adversarial network (GAN) for online data augmentation and introduces a novel regularization technique to stabilize GAN training and prevent mode collapse. Furthermore, our model employs the Random Key method, optimized by the differential evolution (DE) algorithm, to minimize the dependence of the approach on hyperparameters. Our comprehensive testing across diverse datasets, including the center for image and information processing-CSI dataset (CIIP-CSID), the global human image dataset with 10,000 images (GHIM-10 K), and corel 1,000 (Corel-1 K), demonstrates that our model achieves F-measures ranging from 86.758 % to 92.611 %. This underscores the superior performance and versatility of the model in handling varied CSI image classification tasks.