4  Presentation Figures

Author

Benjamin Poulin

Published

July 31, 2025

Figure 4.1 demonstrates the complex operating environment targeted by this project. Shallow water dehazing is unusual in the literature, and its large-particulate, inconsistantly illuminated, high turbidity poses challenges that reduce the effectiveness of current state-of-the-art pipelines.

Figure 4.1: Operating Environment

Figure 4.2 displays sample outputs for pipelines representing major areas of dehazing research. UDCP is model-based, (T)CLAHE is model-free, and EDN-GTM uses machine learning. Parameters for CLAHE & TCLAHE were tuned with a manual grid search on a per-image basis. EDN-GTM is trained on an in-air dataset, so its performance could be improved by retraining on underwater data.

Code 
import numpy as np
import cv2
import os
import matplotlib.pyplot as plt

TILE_WIDTH = 1280
TILE_HEIGHT = 720

OG_IMGS = (
    "/home/mantis/Documents/ms/data/cup/0.png",
    "/home/mantis/Documents/ms/data/turbid/deepblue/prog/1080/13.jpg",
    "/home/mantis/Documents/ms/data/turbid/milk/prog/1080/19.jpg",
    "/home/mantis/Documents/ms/data/lakes/erie_0/0.png",
    "/home/mantis/Documents/ms/data/lakes/erie_2025_03/frames/blue_1.png",
    "/home/mantis/Documents/ms/data/lakes/erie_2025_07/frames/png/LT_35_100_frame.png",
)

SINGLES_DIR = '/home/mantis/Documents/ms/data/output/summer_end/singles/'
ALGO_DIRS = ('hazy', 'udcp', 'clahe', 'tclahe', 'edn-gtm')

IMAGE_NAMES = ('Cup', 'Deepblue', 'Milk', 'Mr. Crab', 'Erie #1', 'Erie #2')

GRID_WIDTH = len(IMAGE_NAMES)
GRID_HEIGHT = len(ALGO_DIRS)

fig, axes = plt.subplots(GRID_HEIGHT, GRID_WIDTH)

for r, dir in enumerate(ALGO_DIRS):
    if dir == 'hazy':
        image_files = OG_IMGS
    else:
        image_files = sorted([(SINGLES_DIR + dir + '/' + f) for f in os.listdir(SINGLES_DIR + dir) if f.lower().endswith(('.png', '.jpg', '.jpeg'))])
    for c, filepath in enumerate(image_files):
        img = cv2.imread(filepath, cv2.IMREAD_COLOR_RGB)
        if img is None:
            raise FileNotFoundError(f'Could not find {filepath}')
        img = cv2.resize(img, (TILE_WIDTH, TILE_HEIGHT), interpolation=cv2.INTER_LINEAR)

        ax = axes[r, c]
        ax.imshow(img)

        ax.set_xticks([])
        ax.set_yticks([])
        ax.set_xticklabels([])
        ax.set_yticklabels([])
        ax.set(
            xlabel=IMAGE_NAMES[c],
            ylabel=ALGO_DIRS[r].upper()
        )
        ax.xaxis.set_label_position('top')
        ax.xaxis.tick_top()
        ax.label_outer()

MARGIN_LR = 0.023
MARGIN_TB = 0.04
fig.subplots_adjust(wspace=0, hspace=0,
                    left=MARGIN_LR, right=1 - MARGIN_LR,
                    top=1 - MARGIN_TB, bottom=MARGIN_TB)
subplot_width = 2
subplot_height = (subplot_width / TILE_WIDTH * TILE_HEIGHT)
fig.set_size_inches(
  (1 + 2 * MARGIN_LR) * subplot_width * GRID_WIDTH,
  (1 + 2 * MARGIN_TB) * subplot_height * GRID_HEIGHT
)
plt.show()
Figure 4.2: Reference metrics

Sample outputs for UDCP (Underwater Dark Channel Prior; model-based), CLAHE (Contrast-Limit Adaptive Histogram Equalization; model-free), TCLAHE (Turbidity-Aware ibid; model-free), and EDN-GTM (Encoder-Decoder Network with Guided Transmission Map; learning). The first row holds the original hazy images. Images are taken from the underwater cup [1] and turbid (deepblue & milk) [2] datasets, as well as two videos of UXOs in lake Erie.