|
Abstract:
|
Convolutional Neural Networks (CNNs) have shown great promise at the task of object detection. Some of its applications, such as security systems, quality control, and autonomous cars, require a high degree of precise detection. Detection performance has implications is security and resource efficiency. Large CNNs are costly to train. Thus, transfer learning has become a common method to cut down on training time and training dataset size requirements. Using transfer learning, we investigated the effects of image preprocessing during the training stage on the robustness of the model's learned features. We investigated this problem in the context of automobile detection. Using a dataset of traffic images from the NYC Department of Transportation, we created a custom dataset for training and evaluating four CNN-based architectures. We applied a number of different transformations to the training set images prior to training, and evaluated the resulting models' precision on the original and transformed images.
This document is based upon work supported by the National Science Foundation under Grant No. 1246818.
|
