| Summary: | Recently, layer-wise learning has been well developed into an alternative training schema of neural networks, aiming to bypass drawbacks brought by traditional backpropagation (BP) learning. A newly error-based forward layer-wise learning method, which is so-called forward progressive learning (FPL), has been used to construct the analytical framework of deep convolutional neural networks (CNNs). The FPL method is capable of more robust learning convergence, better performance and more explainable ability than the well-known stochastic gradient descent (SGD) method. Previous researches related to the FPL method only restrict to the classification task, but the transfer learning abilities of these pre-trained models also need to be investigated to fit into other tasks. In this dissertation project, we proposed a simple object detection architecture, image pyramids and sliding windows (IPSW), to convert pre-trained models into object detectors. Through massive comparisons, it turns out that models pre-trained by the FPL method, especially those subnets in the analytical structure of CNNs, fine-tuned with our proposed IPSW achieve better detection metrics but have less trainable parameters in the pre-training stage than those counterparts with the SGD method. Moreover, we also compared our proposed IPSW with other popular types of object detection architecture, such as R-CNN and Faster R-CNN. Numerical observations indicate that our proposed IPSW is a more suitable option for the evaluation of transfer learning abilities of pre-trained models with the FPL method in the field of object detection.
|
|---|