2  Basic metrics

I evaluated output quality on the DeepBlue reference dataset for TCLAHE, DCP, UDCP, and TCLAHE followed by UDCP. I used reference metrics MSE, PSNR, and SSIM, and referenceless metrics entropy and UCIQE. See Figure 2.1.

import json

PRETTY_METRIC_NAMES = {
  'mse': 'MSE (↓)',
  'psnr': 'PSNR (↑)',
  'ssim': 'SSIM (↑)',
  'entropy': 'Entropy (↑)',
  'uciqe': 'UCIQE (↑)'
}
REF_METRICS = ('mse', 'psnr', 'ssim')
REFLESS_METRICS = ('entropy', 'uciqe')
PIPELINES = ('hazy', 'dcp', 'udcp', 'tclahe', 'tclahe_udcp')
PIPELINE_COLORS = {
  'dcp': 'green',
  'udcp': 'blue',
  'tclahe': 'red',
  'tclahe_udcp': 'purple',
  'hazy': 'black'
}
FILENAMES = ('00gt.png', '01.png', '02.png', '03.png', '04.png', '05.png', '06.png', '07.png', '08.png', '09.png', '10.png', '11.png', '12.png', '13.png', '14.png', '15.png', '16.png', '17.png', '18.png', '19.png', '20.png')
FILENAMES_NO_EXTENSION = [filename[:-4] for filename in FILENAMES]

ref_data = {}
refless_data = {}

def make_metric_data(metrics_json, metric_name: str):
  data = []
  for pipeline in PIPELINES:
    data.append([])
    for filename in FILENAMES:
      data[-1].append(metrics_json[pipeline][filename][metric])

  return data

with open('../data/metrics.json', 'r') as metrics:
  metrics_json = json.load(metrics)
  for metric in REF_METRICS:
    ref_data[metric] = make_metric_data(metrics_json, metric)
  for metric in REFLESS_METRICS:
    refless_data[metric] = make_metric_data(metrics_json, metric)

import matplotlib.pyplot as plt

for metrics, data in ((REF_METRICS, ref_data), (REFLESS_METRICS, refless_data)):
  fig, axs = plt.subplots(nrows=1, ncols=len(metrics))
  axs = axs.flatten()
  for metric, ax in zip(metrics, axs, strict=True):
    ax.grid(alpha=0.7)
    ax.tick_params(axis='x', labelrotation=90)
    ax.set(
      xlabel='File',
      # ylabel=PRETTY_METRIC_NAMES[metric],
      title=PRETTY_METRIC_NAMES[metric]
    )
    for series, label in zip(data[metric], PIPELINES, strict=True):
      ax.plot(
        FILENAMES_NO_EXTENSION, series,
        label=label, color=PIPELINE_COLORS[label]
      )

  handles, labels = axs[0].get_legend_handles_labels()
  fig.legend(handles, labels, loc='upper center', ncol=len(PIPELINES), bbox_to_anchor=(0.5, 1.1))
  fig.set_size_inches(10, 4)
  fig.tight_layout()
  plt.show()

(a) Reference metrics

(b) Referenceless metrics

Figure 2.1: Reference and referenceless Metrics for TCLAHE, DCP, UDCP, and TCLAHE -> UDCP on the DeepBlue dataset. Higher filenames are more turbid. 00gt.png has no turbidity.

See Figure 4.2 for examples of dehazed images.

from pathlib import Path
import cv2

BASE_PATH = Path('/home/mantis/Documents/ms/underwater-dehazing-toolkit/data/deepblue')
FILES = ('02.png', '10.png', '20.png')

fig, axs = plt.subplots(nrows=len(FILES), ncols=len(PIPELINES), sharex=True, sharey=True)
for file, axs_row in zip(FILES, axs, strict=True):
  for pipeline, ax in zip(PIPELINES, axs_row, strict=True):
    image = cv2.imread(str(BASE_PATH / pipeline / file))
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    image_shape = image.shape
    ax.imshow(image)
    ax.set_xticks([])
    ax.set_yticks([])
    ax.set_xticklabels([])
    ax.set_yticklabels([])
    ax.set(
      xlabel=pipeline,
      ylabel=file
    )
    ax.label_outer()
fig.subplots_adjust(wspace=0, hspace=0, left=0, right=1, top=1, bottom=0)
subplot_width = 2
subplot_height = (subplot_width / image_shape[1] * image_shape[0])
fig.set_size_inches(
  subplot_width * 5,
  subplot_height * 3
)
plt.show()

Figure 2.2: Low (02.png), medium (10.png), and high (20.png) turbidity images from the DeepBlue dataset, processed by TCLAHE, DCP, UDCP, and TCLAHE -> UDCP. The hazy column are the original images.

TCLAHE and TCLAHE -> UDCP are consistantly the best performing pipelines (barring their low SSIM scores). DCP seems to modify images the least.

3 Next Steps

• Measure runtimes
• Add learning pipelines
• Choose which Retinex/CLAHE/etc. model-free pipelines to use