Abstract:

In my dissertation, we present multiple instance learning U-net (MILUnet) algorithm and multiple instance learning class activation map (MILCAM) algorithm for weakly supervised semantic segmentation. Both the MILUnet and MILCAM algorithms requires only training images paired with image-level label to classify pixels into one or other classes into images. Compared with supervised semantic segmentation algorithms, our weakly supervised approach relax the requirements of supervisions for semantic segmentation from pixel-level label to image-level label which enormously reduce the labor on constructing the training dataset. We proposed three multiple instance learning (MIL) models to formulate the weakly supervised image segmentation problem as a maximize likelihood problem with respect to both model parameters and pixel-level label of training data. For the MILUnet algorithm, the three models are applied to training images directly used for semantic segmentation. The MILCAM algorithm is a two-step algorithm. The three models are applied to the sampled pixel from the histogram of the training image in the first step. In the first step of MILCAM algorithm, the attention maps are generated by linearly combining the feature maps of training image. Those attention maps from are the used to initialize the pixel-level label of the training data and do semantic segmentation in the second step. Our proposed MILUnet and MILCAM algorithms are applied to do weakly supervised semantic segmentation on the switchgrass minirhizotron dataset and peanut minirhizotron dataset. On both dataset, we achieved good segmentation results.

Links:

Citation:

G. Yu, "Weakly Supervised Image Segmentation with Multiple Instance Learning Neural Network," Ph.D Thesis, Gainesville, FL, 2021.

@phdthesis{Yu2022Thesis,
author = {Guohao Yu},
title = {Weakly Supervised Image Segmentation with Multiple Instance Learning Neural Network},
school = {Univ. of Florida},
year = {2021},
address = {Gainesville, FL},
}