Abstract:
The recent development of deep learning leads to the current surge of breakthroughs in the research field of computer vision. Most deep learning models rely on the convolution operator to extract features. However, the convolution operator has several limitations. For example, convolution is a linear feature extractor based on the correlation between filters and the receptive field. The neural network requires activation functions to achieve non-linearity. Due to the nature of convolution operation, the relative difference between foreground and background information is not guaranteed to be preserved. Instead, convolution operation tends to learn either the foreground or the background information depending on the variance across the classes. Thus, the learned features are usually not specific to the given target. Also, the features learned by convolution have poor interpretability. At last, the convolution filters in each layer are independent of each other. The filters could learn redundant features when the model becomes more complicated. To tackle these issues, the morphological hit-miss transform was proposed to be introduced in deep neural networks as a different type of feature extractor to learn morphological features from the given data, as they offer benefits such as interpretability and the ability to analyze and process geometric structures. In addition, the hit-or-miss transform in mathematical morphology considers both foreground and background information when evaluating shapes in images, providing more constrained features to help handle variations. In this work, the properties of morphological hit-miss transform in the deep neural network was studied, and compared with convolution operation to further demonstrate the similarities and difference between these two feature extractors. Finally, two novel models were developed to combine convolution and morphological hit-miss operations to enrich feature space in the deep neural network. Experiments were designed to evaluate the proposed approaches on synthetic datasets and benchmark datasets to show the improvement of model performance by taking advantage of both operators.
Links:
Citation:
W. Xu, “Deep Morph-Convolutional Neural Network: Combining Morphological Transform and Convolution in Deep Neural Networks.” University of Florida, 2023.@book{xu2023deep,
title={Deep Morph-Convolutional Neural Network: Combining Morphological Transform and Convolution in Deep Neural Networks},
author={Xu, Weihuang},
year={2023},
publisher={University of Florida}
}