Enhancing Accuracy in Detecting Objects in Video Surveillance Systems using Transfer Leaning
Senthil Pandi S, S. Shamili Shanmugapriya, S. Chakaravarthi, P Kumar · 2024
In surveillance imaging, where huge volumes of complicated data contain the secret to discovering new surveillance insights, some impact is most felt. Deep learning has enormous promise for improving surveillance imaging. The prognosis for public can be greatly enhanced with the early and precise detection of disorders such as object detection. One of the significant problems in Convolutional Neural Networks is the tendency to overfit and perform poorly in generalizing on data not seen before, especially small datasets. Transfer learning using deep CNNs is of particular benefit in overcoming this problem. Transfer learning will leverage a model already trained, improving the efficiency of training and reducing the potential risk of overfitting. This approach is advisable when training models on limited datasets. This research Comparing the performance of four approaches-CNN, GAN, RNN, and Transfer Learning-on three datasets-KITTI, COCO, and Cityscapes-on three metrics-accuracy, recall, and F1-Score. The graph reads that transfer learning performs better on all measures across all datasets.