去重 CV 检测框的 50 pix 像素值内的重复框:
如果缺陷类型一样,有重复框则取检测分数最大的框;
如果缺陷类型不一致,则保留两个框;
思想:
- 先找到每一个缺陷的几何中心 50 pix 内的所有框;
- 按照框进行缺陷类型分组:做去重处理;
- 删除位于同一范围内的缺陷;
- 合并小于 50 像素值和大于 50 pix 的框;
用法:依赖以下 3 个方法,直接调用第 3 个方法 remove_duplicate(txt_path, nearby_pix)
输出是去重后的生成 txt 。原始 txt 格式如下:
id 名称,x, y,w, h, 检测分数,缺陷类型,时间
# txt to df and select the max score
def choose_only_one(nearby_list, defect_list_):
df_nearby = pd.DataFrame()
id_list = []
x_list = []
y_list = []
w_list = []
h_list = []
score_list = []
defect_list = []
time_list = []
for ele_nearby in nearby_list:
ele_nearby = ele_nearby.split(',')
id_list.append(ele_nearby[0])
x_list.append(ele_nearby[1])
y_list.append(ele_nearby[2])
w_list.append(ele_nearby[3])
h_list.append(ele_nearby[4])
score_list.append(ele_nearby[5])
defect_list.append(ele_nearby[6])
time_list.append(ele_nearby[7][:-1])
# time_list.append(ele_nearby[7])
df_nearby['id'] = id_list
df_nearby['x'] = x_list
df_nearby['y'] = y_list
df_nearby['w'] = w_list
df_nearby['h'] = h_list
df_nearby['score'] = score_list
df_nearby['defect'] = defect_list
df_nearby['date'] = time_list
result_list = []
df_result = pd.DataFrame()
for ele_defect in defect_list_:
df_nearby_temp = df_nearby[df_nearby['defect']==str(ele_defect)]
df_nearby_max = df_nearby_temp[df_nearby_temp['score'] == np.max(df_nearby_temp['score'])]
# result_list.append(df_nearby_max)
df_result = pd.concat([df_result, df_nearby_max], axis=0)
return df_result
def txt_to_df(nearby_list):
df_nearby = pd.DataFrame()
id_list = []
x_list = []
y_list = []
w_list = []
h_list = []
score_list = []
defect_list = []
time_list = []
for ele_nearby in nearby_list:
ele_nearby = ele_nearby.split(',')
id_list.append(ele_nearby[0])
x_list.append(ele_nearby[1])
y_list.append(ele_nearby[2])
w_list.append(ele_nearby[3])
h_list.append(ele_nearby[4])
score_list.append(ele_nearby[5])
defect_list.append(ele_nearby[6])
time_list.append(ele_nearby[7][:-1])
# time_list.append(ele_nearby[7])
df_nearby['id'] = id_list
df_nearby['x'] = x_list
df_nearby['y'] = y_list
df_nearby['w'] = w_list
df_nearby['h'] = h_list
df_nearby['score'] = score_list
df_nearby['defect'] = defect_list
df_nearby['date'] = time_list
return df_nearby
# remove dumplicate jiaoji txt
def remove_duplicate(txt_path, nearby_pix):
'''
txt_path : txt 文件的路径
nearby_pix : 邻近范围的像素值
'''
work_dir = './factory_result_dashboard_txt/'
if(not os.path.exists(work_dir)):
os.makedirs(work_dir)
f1=open(txt_path,'r')
f1_1 = open(txt_path, 'r').read()
if(len(f1_1)<=10):
if(os.path.exists(txt_path)):
shutil.copy(txt_path, work_dir)
lines_1_list = list(set(f1.readlines()))
lines_2_list = copy.deepcopy(lines_1_list)
single_defect_list = []
df_single_defect = pd.DataFrame()
# single_defect_list_1 = []
for line_1 in lines_1_list:
line_1_ele = line_1.split(',')
line_1_x = float(line_1_ele[1])
line_1_y = float(line_1_ele[2])
line_1_w = float(line_1_ele[3])
line_1_h = float(line_1_ele[4])
line_1_score = float(line_1_ele[5])
line_1_defect = float(line_1_ele[6])
line_1_centre_x = line_1_x + line_1_w // 2
line_1_centre_y = line_1_y + line_1_h // 2
nearby_list = []
defect_list = []
for line_2 in lines_2_list:
line_2_ele = line_2.split(',')
line_2_x = float(line_2_ele[1])
line_2_y = float(line_2_ele[2])
line_2_w = float(line_2_ele[3])
line_2_h = float(line_2_ele[4])
line_2_score = float(line_2_ele[5])
line_2_defect = float(line_2_ele[6])
line_2_centre_x = line_2_x + line_2_w // 2
line_2_centre_y = line_2_y + line_2_h // 2
if(line_1==line_2):
continue
# if(line_1 in single_defect_list):
# break
dis = np.sqrt(math.pow(abs(line_1_centre_x - line_2_centre_x), 2) + math.pow(abs(line_1_centre_y - line_2_centre_y), 2))
nearby_pix = float(nearby_pix)
if(dis <= nearby_pix):
if(line_1 not in nearby_list):
nearby_list.append(line_1)
defect_list.append(line_1_defect)
if(line_2 not in nearby_list):
nearby_list.append(line_2)
defect_list.append(line_2_defect)
defect_list = list(set(defect_list))
defect_list = map(int, defect_list)
df_res = choose_only_one(nearby_list, defect_list) # select only one defect
# single_defect_list.extend(df_res)
df_single_defect = pd.concat([df_single_defect, df_res], axis=0)
nearby_list = map(str, nearby_list)
for ele_nearby in nearby_list:
lines_1_list.remove(ele_nearby)
df_lines_1 = txt_to_df(lines_1_list)
df_single_defect = pd.concat([df_single_defect, df_lines_1], axis=0)
df_single_defect = df_single_defect.drop_duplicates()
# df_single_defect = df_single_defect.drop([0])
df_single_defect.to_csv(work_dir + txt_path, sep=',', header=None, index=False)