11. Improving object detection performance on Small-Scale datasets through Pre-Trained deep feature extraction

Tóm tắt

Transfer learning has emerged as an effective approach for improving object recognition performance, particularly in scenarios with limited training data. In this study, a model based on the VGG16 architecture was developed by exploiting pre-trained deep features to address the object recognition problem on a custom dataset consisting of 16 categories of common objects. The proposed model was fine-tuned using weights pre-trained on the ImageNet dataset and evaluated through five independent training runs to verify the stability of the obtained results. Experimental results demonstrated that the model achieved an average accuracy of 90.56% with a standard deviation of 2.15%, indicating high recognition performance, strong stability, and clear advantages over training-from-scratch approaches. In addition, the proposed method achieved a mean Average Precision (mAP) of 74.8% at IoU = 0.5 and 59.3% at IoU = 0.75, demonstrating robust object detection capability under different evaluation thresholds. Notably, the detection performance for small objects was improved (AP_small = 38.2%), confirming the effectiveness of transfer learning in extracting fine-grained features. However, the model still encountered challenges in accurately detecting extremely small or occluded objects, highlighting limitations in feature representation at low scales. Overall, this study confirms the effectiveness of transfer learning on academic-scale datasets and provides experimental evidence for its practical applicability. Future work will focus on integrating multi-scale feature representations and exploring more advanced architectures to further improve detection performance.