def extract_number(digit_string):
#function to get the output from tesseract and the output will be converted to float
try:
pred = float(digit_string.replace(" ", ""))
except ValueError:
scatteract_logger.get_logger().error("Output from tesseract is not a float " + digit_string)
pred = None
return pred
def extract_number(digit_string):
"""
Method used to clean up the output from tesseract.
Inputs:
digit_string (string): String output from tesseract.
Outputs:
pred (float) : If the input was a number, this is the float version of this number,
else the output is None.
"""
try:
pred = float(digit_string.replace(" ", ""))
except ValueError:
scatteract_logger.get_logger().error(
"Output from tesseract is not a float " + digit_string)
pred = None
return pred
def get_metrics(self, df_dict_pred, coord_idl, csv_output_dir = None, max_dist_perc = 2.0, quick = False):
"""
Method which computes the precision and recall for all the plots.
Inputs:
df_dict_pred (dictionary): Dictionnary of dataframes wich contains the predicted points in label coordinates.
coord_idl (string): Path to the idl files which contain the ground truth coordinates
max_dist_perc (float): Maximum tolerated percentage difference between prediction and ground truth.
quick (boolean): Boolean to decide weather to approximate the precision recall quantities.
Outputs:
df_prec_recall : (pandas dataframe) Pandas dataframe of precisions and recalls for each plot.
"""
df_dict_true = parse_coords(coord_idl)
precision_list, recall_list, image_name_list = [], [], []
count_good = 0
count_perfect = 0
count_bad = 0
metrics_logger = scatteract_logger.get_logger()
for file_name in df_dict_pred:
df_pred = df_dict_pred[file_name]
df_true = df_dict_true.get(file_name,None)
if df_true is not None:
prec, rec = self.get_precision_recall(df_pred,df_true,
max_dist_perc = max_dist_perc,
norm = (df_true.max(axis=0)-df_true.min(axis=0)), quick = quick)
precision_list.append(prec)
recall_list.append(rec)
image_name_list.append(file_name)
if rec>=0.8 and prec>=0.8:
count_good+=1
if rec==1.0 and prec==1.0:
count_perfect+=1
if rec<=0.1 and prec<=0.1:
count_bad+=1
else:
metrics_logger.warn("No ground truth for :" + file_name)
metrics_logger.info("Percentage of good extraction (recall and precision above 80%): {}".format(float(count_good)/len(precision_list)))
metrics_logger.info("Percentage of perfect extraction (recall and precision at 100%): {}".format(float(count_perfect)/len(precision_list)))
metrics_logger.info("Percentage of bad extraction (recall and precision below 10%): {}".format(float(count_bad)/len(precision_list)))
metrics_logger.info("Precision: {}".format(np.mean(precision_list)))
metrics_logger.info("Recall: {}".format(np.mean(recall_list)))
metrics_logger.info("F1 score: {}".format(2*np.mean(recall_list)*np.mean(precision_list)/(np.mean(recall_list)+np.mean(precision_list))))
df_prec_recall = pd.DataFrame({"image_name":image_name_list,"recall":recall_list,"precision":precision_list})
if csv_output_dir is not None:
df_prec_recall.to_csv(csv_output_dir + "/" + "precision_recall_list.csv")
metrics_logger.debug("Saving a csv of precisions and recalls : {}".format(csv_output_dir + "/" + "precision_recall_list.csv"))
return df_prec_recall
Example of command-line usage:
TEST: (requires a dict of ground truth for ticks, labels and points bounding boxes, and an idl file for the ground truth coordinate)
python scatter_extract.py --model_dict '{"ticks":"./output/lstm_rezoom_plot_ticks_2017_04_11_19.52", "labels":"./output/lstm_rezoom_plot_labels_2017_04_11_01.14","points":"./output/lstm_rezoom_plot_points_2017_04_14_15.58"}' \
--iteration 125000 --image_dir data/plot_test/ --true_idl_dict '{"ticks":"./data/plot_test/ticks.idl","labels":"./data/plot_test/labels.idl", "points":"./data/plot_test/points.idl"}' \
--image_output_dir image_output --csv_output_dir csv_output --true_coord_idl ./data/plot_test/coords.idl
PREDICT: (requires an idl file which has a list of images to test on)
python scatter_extract.py --model_dict '{"ticks":"./output/lstm_rezoom_plot_ticks_2017_04_11_19.52", "labels":"./output/lstm_rezoom_plot_labels_2017_04_11_01.14","points":"./output/lstm_rezoom_plot_points_2017_04_14_15.58"}' \
--iteration 125000 --image_dir data/plots_real/ --predict_idl ./data/plots_real/test_real.idl --image_output_dir image_output --csv_output_dir csv_output
"""
mylogger = scatteract_logger.get_logger()
parser = argparse.ArgumentParser()
parser.add_argument('--model_dict', help='Directory for the object detection models', required=True, type=json.loads)
parser.add_argument('--iteration', help='Iteration number for the trained models', required=True)
parser.add_argument('--image_dir', help='Directory of the images', required=True)
parser.add_argument('--true_idl_dict', help='Path of the ground truth idls', required=False, type=json.loads, default=None)
parser.add_argument('--predict_idl', help='Path of an idl file which list the images to predict on', required=False, default=None)
parser.add_argument('--image_output_dir', help='Directory to output images with bounding boxes', required=True)
parser.add_argument('--csv_output_dir', help='Directory to output csv of results', required=True)
parser.add_argument('--true_coord_idl', help='Idl of the ground truth coordinates', required=False, default=None)
parser.add_argument('--conf_threshold', help='Confidence threshold', required=False, default=0.3)
parser.add_argument('--max_dist_perc', help='Maximum percent distance to be considered true positive', required=False, default=2.0)
args = vars(parser.parse_args())
def get_metrics(self,
df_dict_pred,
coord_idl,
csv_output_dir=None,
max_dist_perc=2.0,
quick=False):
# function for saving the Precision and recall in csv file.
df_dict_true = read_coordinates(coord_idl)
precision_list, recall_list, image_name_list = [], [], []
count_good = 0
count_perfect = 0
count_bad = 0
metrics_logger = scatteract_logger.get_logger()
for name_f in df_dict_pred:
df_pred = df_dict_pred[name_f]
df_true = df_dict_true.get(name_f, None)
if df_true is not None:
prec, rec = self.get_precision_recall(
df_pred,
df_true,
max_dist_perc=max_dist_perc,
norm=(df_true.max(axis=0) - df_true.min(axis=0)),
quick=quick)
precision_list.append(prec)
recall_list.append(rec)
image_name_list.append(name_f)
if rec >= 0.8 and prec >= 0.8:
count_good += 1
if rec == 1.0 and prec == 1.0:
count_perfect += 1
if rec <= 0.1 and prec <= 0.1:
count_bad += 1
else:
metrics_logger.warn("No ground truth for :" + name_f)
metrics_logger.info(
"Percentage of good extraction (recall and precision above 80%): {}"
.format(float(count_good) / len(precision_list)))
metrics_logger.info(
"Percentage of perfect extraction (recall and precision at 100%): {}"
.format(float(count_perfect) / len(precision_list)))
metrics_logger.info(
"Percentage of bad extraction (recall and precision below 10%): {}"
.format(float(count_bad) / len(precision_list)))
metrics_logger.info("Precision: {}".format(np.mean(precision_list)))
metrics_logger.info("Recall: {}".format(np.mean(recall_list)))
metrics_logger.info("F1 score: {}".format(
2 * np.mean(recall_list) * np.mean(precision_list) /
(np.mean(recall_list) + np.mean(precision_list))))
df_prec_recall = pd.DataFrame({
"image_name": image_name_list,
"recall": recall_list,
"precision": precision_list
})
if csv_output_dir is not None:
df_prec_recall.to_csv(csv_output_dir + "/" +
"precision_recall_list.csv")
metrics_logger.debug(
"Saving a csv of precisions and recalls : {}".format(
csv_output_dir + "/" + "precision_recall_list.csv"))
return df_prec_recall
import pyocr
import pyocr.builders
import pandas as pd
import numpy as np
import cv2
from scipy import ndimage
import time
import argparse
import scatteract_logger
tools = pyocr.get_available_tools()
if len(tools) == 0:
print("No OCR tool found")
sys.exit(1)
tool = tools[0]
scatteract_logger.get_logger().info("Will use tool '%s'" % (tool.get_name()))
def compute_skew(image):
"""
Method which computes the skew of a label.
Inputs:
image (numpy array): Numpy array of the image label.
Output:
angle (float) Angle needed to unskew the label.
contours_len: Number of independent object detected
(if the image is clean, this is the number of digitsminus signs, and points)
"""
image = ImageOps.invert(image)
