Computer vision object detection and image recognition algorithm optimization based on self-supervised learning

Traditional target detection and image recognition algorithms are highly dependent on data labeling, By leveraging the inherent information within data to generate supervisory signals, self-supervised learning lessens the reliance on labeled datasets and enhances the model's capacity to generalize. Therefore, this paper designs three self-supervised learning tasks: image rotation angle prediction, image block position prediction and image color reconstruction to help the model learn useful features such as orientation, spatial layout and color characteristics. A vast collection of unlabeled images is first utilized for self-supervised learning, laying the foundation through pre-training. This pre-trained model is then adapted and fine-tuned for downstream tasks such as object detection and image classification. To enhance feature integration, an attention-based fusion module is introduced, effectively merging the learned representations from pre-training with task-specific features. Additionally, a combined loss function is formulated to further refine and boost the model’s overall performance. The experiment uses ImageNet-1k unlabeled data set for pre-training, COCO 2017 data set for target detection and evaluation, and PASCAL VOC 2012 data set for image recognition and evaluation. The results show that multi-task self-supervised learning significantly improves the performance of the model. SSL-Multi+AF method combined with attention feature fusion has achieved a mAP of 44.8 in target detection tasks, which is 6.6 higher than Baseline, and has excellent performance on targets of different scales. In the image recognition task, it also achieved high accuracy of Top-1 and Top-5. Self-supervised learning performs well in low labeled data, and the model performance is improved by 46.8% when only 10% labeled data is used, which effectively alleviates the dependence on labeled data. This study introduces a fresh perspective and technique to enhance algorithms for object detection and image recognition in computer vision.