We propose the implementation of transfer learning from natural images to audio-based images using self-supervised learning schemes. Through self-supervised learning, convolutional neural networks (CNNs) can learn the general representation of natural images without labels. In this study, a convolutional neural network was pre-trained with natural images (ImageNet) via self-supervised learning; subsequently, it was fine-tuned on the target audio samples. Pre-training with the self-supervised learning scheme significantly improved the sound classification performance when validated on the following benchmarks: ESC-50, UrbanSound8k, and GTZAN. The network pre-trained via self-supervised learning achieved a similar level of accuracy as those pre-trained using a supervised method that require labels. Therefore, we demonstrated that transfer learning from natural images contributes to improvements in audio-related tasks, and self-supervised learning with natural images is adequate for pre-training scheme in terms of simplicity and effectiveness.

Building training datasets for perception systems in autonomous driving and industrial inspection requires large volumes of pixel-level annotations, and the annotation cost increases with scene complexity and the number of object categories. Existing annotation platforms combine foundation models with active learning, and some offer confidence-based filtering. However, existing systems do not account for the class imbalance that arises during iterative label pool expansion, nor study how semi-supervised learning (SSL), calibration, and selective review interact across annotation rounds. We propose an annotation system that reduces the manual annotation effort through calibrated confidence scores: predictions below a minimum threshold are discarded, high-confidence predictions are accepted without human review, and the remaining uncertain predictions are flagged for manual inspection. A controlled annotation experiment on Cityscapes shows that the system reduces human annotation time by 1.9x and improves Average Precision (AP) by 1.8 points over an active-learning baseline with Mask2Former and SAM 2, while auto-accepting 74.3% of predictions with 85.8% precision at the final round. An offline ablation shows that SSL and calibration jointly reduce false positives by 40.4%, with auto-accept precision improving across rounds rather than declining under supervised-only training. Compared to unassisted manual polygon labeling, the proposed system reduces human annotation time by 3.2x while improving AP by 4.5 points under the same 150-image budget.