Deep Learning Models of Artificial Intelligence in Computer Vision

Deep learning models in computer vision face challenges such as high computational resource demands and limited generalization in practical scenarios. To address these issues, this study proposes a systematic optimization framework integrating transfer learning, self-supervised learning, model pruning, and data augmentation. Specifically, transfer learning leverages pre-trained models for efficient feature extraction, while self-supervised learning reduces dependency on labeled data through contrastive pre-training. Model pruning and quantization streamline network complexity, and advanced data augmentation (e.g., Mixup, CutMix) enhances robustness. Experiments across image classification, object detection, and semantic segmentation tasks demonstrate significant improvements: ResNet50 achieved 79.2% classification accuracy, pruned Faster R-CNN reduced parameters by 30% with minimal accuracy loss, and combined self-supervised learning with augmentation boosted semantic segmentation mIoU to 73.4%. These strategies effectively balance accuracy, efficiency, and adaptability, offering practical solutions for resource-constrained applications.