I am building an image-processing app in dash, where the user uploads an image, selects a smoothing filter and then chooses between k-means clustering and thresholding to identify the foreground object from the background.
As shown below, I have a callback with a filters-drop down (which works fine as both filters take the same type of input - in this case the kernel-height-slider and the kernel-width-slider.
import numpy as np
import cv2 as cv
from dash import Dash, dcc, html
from dash.dependencies import Input, Output
import dash_daq as daq
import io
import os
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from PIL import Image
import base64
import plotly.express as px
import plotly.graph_objects as go
external_stylesheets = ['https://codepen.io/chriddyp/pen/bWLwgP.css']
app = Dash(__name__, external_stylesheets=external_stylesheets, suppress_callback_exceptions=True)
img_path = "assets/lyco_WOL_00140.tif"
def preprocessing(img_path):
# Image processing
img = cv.imread(img_path)
imgYCC = cv.cvtColor(img, cv.COLOR_BGR2YCR_CB)
gray = imgYCC[:,:,0]
return gray
preprocessed_img = preprocessing(img_path)
filters = {
"gaussian": lambda img, kh, kw: cv.GaussianBlur(img, (kh,kw), 0),
"2dconv": lambda img, kh, kw: cv.filter2D(img, -1, np.ones((kh,kw), np.float32)/25)
}
def otsu_thresholding(img, min_thr, max_thr):
return cv.threshold(img, min_thr, max_thr, cv.THRESH_OTSU)
def kmeans_clustering(img, K, n_iter, accuracy):
Z = img.reshape((-1,2))
Z = np.float32(Z)
criteria = (cv.TERM_CRITERIA_EPS + cv.TERM_CRITERIA_MAX_ITER, n_iter, accuracy)
_, label, center = cv.kmeans(Z, K, None, criteria, n_iter, cv.KMEANS_RANDOM_CENTERS)
center = np.uint8(center)
res = center[label.flatten()]
res2 = res.reshape((img.shape))
return res2
detectors = {
"otsu": otsu_thresholding,
"kmeans": kmeans_clustering
}
@app.callback(
Output('smoothed-img','src'),
[Input('filters-dropdown', 'value'),
Input('kernel-height-slider', 'value'),
Input('kernel-width-slider', 'value'),
Input('detectors-radio', 'value'),
Input('otsu-min-threshold', 'value'),
Input('otsu-max-threshold', 'value'),
Input('kmeans-kvalue-input', 'value'),
Input('kmeans-niter-input', 'value'),
Input('kmeans-accuracy-input', 'value'),
]
)
# from PreprocessingClass import PreprocessClass
def update_overlay_base64(selected_filter, kernel_height, kernel_width,
selected_detector, otsu_min_thr, otsu_max_thr,
kmeans_k, kmeans_niter, kmeans_accuracy):
if not selected_filter:
return None
# prepro_class = PreprocessClass(selected_filter = selected_filter)
# if selected_filter == "kmeans":
# prepro_class.kmeans_clustering(img, K, n_iter, accuracy)
filtered_img = filters[selected_filter](preprocessed_img, kernel_height, kernel_width)
if selected_detector == "otsu" and otsu_min_thr is not None and otsu_max_thr is not None:
processed_img = otsu_thresholding(filtered_img, otsu_min_thr, otsu_max_thr)[1]
elif selected_detector == "kmeans" and None not in (kmeans_k, kmeans_niter, kmeans_accuracy):
processed_img = kmeans_clustering(filtered_img, kmeans_k, kmeans_niter, kmeans_accuracy)
edge_img = cv.Canny(processed_img, 10, 150)
edge_img = np.array(edge_img, dtype='float64')
edge_img[edge_img==0] = np.nan
rows, cols = np.where(edge_img == 255)
min_row, max_row = rows.min(), rows.max()
min_col, max_col = cols.min(), cols.max()
# Plot with matplotlib
fig, ax = plt.subplots()
ax.imshow(processed_img, cmap='Spectral_r')
ax.imshow(edge_img, cmap='autumn', alpha=0.5)
plt.gca().add_patch(
plt.Rectangle((min_col-1, min_row-1),
max_col - min_col + 2,
max_row - min_row + 2,
edgecolor='cyan',
facecolor='none',
linewidth=1)
)
ax.axis('off')
# Save plot to buffer
buf = io.BytesIO()
plt.savefig(buf, format='png', bbox_inches='tight', pad_inches=0)
plt.close(fig)
buf.seek(0)
# Encode to base64
encoded = base64.b64encode(buf.read()).decode()
return f"data:image/png;base64,{encoded}"
# pil_img = Image.fromarray(edge_otsu_img)
# buffer = io.BytesIO()
# pil_img.save(buffer, format="PNG")
# encoded_image = base64.b64encode(buffer.getvalue()).decode()
# Dash app layout
app.layout = html.Div(
children=[
html.H1('Test Flame Analyser v0.3'),
html.Div(
children=[
dcc.Dropdown(
id='filters-dropdown',
options = [{
'label':'Gaussian', 'value':'gaussian'},
{'label':'2D Convolution', 'value':'2dconv'}]
),
html.Img(id='smoothed-img', style={"maxWidth": "100%"}),
daq.Slider(id='kernel-height-slider',
min=1, max=25, value=5,
marks = {'5':'5', '15':'15', '25':'25'},
handleLabel={"showCurrentValue": True,"label": "Height"},
step=2
),
daq.Slider(id='kernel-width-slider',
min=1, max=25, value=5,
marks = {'5':'5', '15':'15' ,'25':'25'},
handleLabel={"showCurrentValue": True,"label": "Width"},
step=2
),
dcc.RadioItems(
id='detectors-radio',
options=[
{'label': 'Otsu Thresholding', 'value': 'otsu'},
{'label': 'K-means Clustering', 'value': 'kmeans'}
],
value='otsu'
),
html.Div(id='detector-params'),
html.Div(id='output')
])
# dcc.Graph(id='figure', figure=fig),
])
@app.callback(
Output('detector-params', 'children'),
[Input('detectors-radio', 'value')]
)
def render_detector_params(detector_name):
if detector_name == 'otsu':
return html.Div([
dcc.Input(id='otsu-min-threshold', type='number', value=0, placeholder="Min Threshold"),
dcc.Input(id='otsu-max-threshold', type='number', value=255, placeholder="Max Threshold"),
])
elif detector_name == 'kmeans':
return html.Div([
#dcc.Dropdown(id='kmeans-kvalue-dropdown', options=[{'label':'2', 'value':2}, {'label':'3', 'value':3}], value=2),
dcc.Input(id='kmeans-kvalue-input', type='number', value=2, placeholder="K (# clusters)"),
dcc.Input(id='kmeans-niter-input', type='number', value=10, placeholder="Max Iterations"),
dcc.Input(id='kmeans-accuracy-input', type='number', value=1.0, placeholder="Accuracy (epsilon)", step=0.01),
])
# def update_figure(img):
# fig = px.imshow(img)
# fig.show()
# return fig
if __name__ == "__main__":
app.run(debug=True)
However, when choosing from the detectors-dropdown, depending on which one is selected the inputs for the image processing function will either be otsu-min-threshold and otsu-max-threshold (if otsu is selected) OR kmeans-kvalue-input, kmeans-niter-input and kmeans-accuracy-input (if kmeans is selected).
The issue is that the callback needs all the Inputs to work properly and throws errors as inevitably, one set of input is not available when the other option is selected. As a result I get the error “A nonexistent object was used in an Input of a Dash callback”. Is there a way to work around this? Can I do conditional callbacks or something similar?
I found something on this plotly forum page which I am about to try but not sure if it will work.
Thank you!
