def generate_img_xml(img_save_name, imgw, imgh, part_imgw, part_imgh,
transforms, imgs_dir, xmls_dir):
'''生成拼接图和拼接xml
Args:
img_save_name:
imgw,imgh:生成总图宽高
transforms:转换操作
imgs_dir:图源目录
xmls_dir:图源对应的xml目录
return:
img:总图
createdxml：总图创建的xml,ET.parse()
'''
createdxml = CreateXML(img_save_name, imgw, imgh, 3)
img = np.zeros((imgh, imgw, 3), dtype=np.uint8)
part_cols = int(imgw / part_imgw)
part_rows = int(imgh / part_imgh)
for row in range(part_rows):
for col in range(part_cols):
start_row = row * part_imgh
start_col = col * part_imgw
part_img_file = utils.randomChoiceIn(imgs_dir)
part_img = cv2.imread(os.path.join(imgs_dir, part_img_file))
part_xml_file = os.path.join(xmls_dir,
part_img_file.split('.')[0] + '.xml')
part_xml_tree = voc_xml.get_xml_tree(part_xml_file)
img, createdxml = mosaic_img_xml(img, part_img, createdxml,
part_xml_tree, transforms,
start_row, start_col)
return img, createdxml
def rotImg_xml_centre_from_dirs(imgs_dir,bk_imgs_dir,xmls_dir,rot_img_save_dir,rot_xmls_save_dir,img_suffix,
name_suffix,rot_angles,randomAngleRange=[0,360],random_num=1,randomRotation=False,scale=1.0,correction = True):
'''
旋转指定路径下的所有图片和xml,以每张图片中心点为旋转中心，并存储到指定路径
Args:
imgs_dir,bk_imgs_dir,xmls_dir: 待旋转图片、背景图片、原始xml文件存储路径
rot_img_save_dir,rot_xmls_save_dir：旋转完成的图片、xml文件存储路径
img_suffix: 图片可能的后缀名['.jpg','.png','.bmp',..]
name_suffix：旋转完成的图片、xml的命名后缀标识
rot_angles: 指定旋转角度[ang1,ang2,ang3,...]
randomAngleRange: 随机旋转上下限角度[bottom_angle,top_angle]
random_num: 随机旋转角度个数，randomRotation=True时生效
randomRotation: 使能随机旋转
scale: 放缩尺度
correction: bool,修正旋转后的目标框为正常左上右下坐标
'''
for root,dirs,files in os.walk(xmls_dir):
for xml_name in files:
xml_file = os.path.join(xmls_dir,xml_name)
img_file = None
for suffix in img_suffix:
#print(os.path.join(imgs_dir,xml_name.split('.')[0]+suffix))
if os.path.exists(os.path.join(imgs_dir,xml_name.split('.')[0]+suffix)):
img_file = os.path.join(imgs_dir,xml_name.split('.')[0]+suffix)
break
if img_file is None:
print("there has no image for ",xml_name)
continue
img = cv2.imread(img_file)
imgh,imgw,n_channels = img.shape
rot_num = random_num
if not randomRotation:
rot_num = len(rot_angles)
for i in range(rot_num):
r_angle = 0
if randomRotation:
r_angle = random.randint(randomAngleRange[0],randomAngleRange[1])
else:
r_angle = rot_angles[i]
bk_img = cv2.imread(os.path.join(bk_imgs_dir,utils.randomChoiceIn(bk_imgs_dir)))
rot_img_name = xml_name.split('.')[0]+'_'+name_suffix+str(r_angle)+'.'+img_file.split('.')[-1]
imgRotation,xmlRotation=generate_rotImg_xml(img,bk_img,voc_xml.get_xml_tree(xml_file),[int(imgw/2),int(imgh/2)],rot_img_name,r_angle,scale,correction)
cv2.imwrite(os.path.join(rot_img_save_dir,rot_img_name),imgRotation)
xmlRotation.save_xml(rot_xmls_save_dir,rot_img_name.split('.')[0]+'.xml')
def transform_img(src_img, transforms):
'''图像变换
Args:
src_img:源图片
transforms:转换操作[{'opt':'rotate','cterxy':[],'imgwh':[],'rot_angle':0,'randomRotation':False,\
'randomAngleRange':[0,360],'scale':1.0,'correction':True,'bk_imgs_dir':'xxx'},
{'opt':'crop','crop_type':RANDOM_CROP,'dsize':(0,0),'top_left_x':0,'top_left_y':0,'fw':0.5,'fh':0.7,'random_wh':False ,'iou_thr':0.5},
{'opt':'flip','flip_type':-1,'random_flip':True,'imgwh':[]},
{'opt':'resize','fx':0.5,'fy':0.5,'dsize':(0,0),'imgwh':[]}]
return：
transformed_img:变换后的图片
certain_transforms:实际变换操作参数
'''
certain_transforms = copy.deepcopy(transforms)
imgh, imgw, depth = src_img.shape
imgwh = [imgw, imgh]
transformed_img = src_img
for operate in certain_transforms:
operate['imgwh'] = imgwh #每一种操作的输入图片宽高
if 'rotate' == operate['opt']:
bk_img = cv2.imread(
os.path.join(operate['bk_imgs_dir'],
utils.randomChoiceIn(operate['bk_imgs_dir'])))
cterxy = [int(imgw / 2), int(imgh / 2)]
rot_angle = operate['rot_angle']
if operate['randomRotation']:
rot_angle = random.randint(operate['randomAngleRange'][0],
operate['randomAngleRange'][1])
transformed_img = rotate.rot_img_and_padding(
transformed_img, bk_img, cterxy, rot_angle, operate['scale'])
operate['cterxy'] = cterxy
operate['rot_angle'] = rot_angle
elif 'resize' == operate['opt']:
resize_imgw, resize_imgh = imgwh[0], imgwh[1]
if (0, 0) == operate['dsize']:
resize_imgw = imgw * operate['fx']
resize_imgh = imgh * operate['fy']
else:
resize_imgw, resize_imgh = operate['dsize']
transformed_img = resize.resize_img(transformed_img,
operate['dsize'],
operate['fx'], operate['fy'])
imgwh = [resize_imgw, resize_imgh]
operate['fx'] = resize_imgw / operate['imgwh'][0]
operate['fy'] = resize_imgh / operate['imgwh'][1]
elif 'crop' == operate['opt']:
crop_imgw, crop_imgh = operate['dsize']
if (0, 0) == operate['dsize'] and not operate['random_wh']:
crop_imgw = int(operate['imgwh'][0] * operate['fw'])
crop_imgh = int(operate['imgwh'][1] * operate['fh'])
elif operate['random_wh']:
crop_imgw = int(operate['imgwh'][0] *
(operate['fw'] + random.random() *
(1 - operate['fw'])))
crop_imgh = int(operate['imgwh'][1] *
(operate['fh'] + random.random() *
(1 - operate['fh'])))
if 'CENTER_CROP' == operate['crop_type']:
top_left_x, top_left_y = int(operate['imgwh'][0] / 2 -
crop_imgw /
2), int(operate['imgwh'][1] / 2 -
crop_imgh / 2)
elif 'RANDOM_CROP' == operate['crop_type']:
top_left_x, top_left_y = random.randint(
0, operate['imgwh'][0] - crop_imgw - 1), random.randint(
0, operate['imgwh'][1] - crop_imgh - 1)
else:
top_left_x, top_left_y = operate['top_left_x'], operate[
'top_left_y']
transformed_img = crop.crop_img(transformed_img, top_left_x,
top_left_y, crop_imgw, crop_imgh)
imgwh = [crop_imgw, crop_imgh]
operate['top_left_x'], operate[
'top_left_y'] = top_left_x, top_left_y
operate['crop_w'], operate['crop_h'] = crop_imgw, crop_imgh
elif 'flip' == operate['opt']:
flip_type = operate['flip_type']
if operate['random_flip']:
flip_type = random.randint(-1, 1)
transformed_img = flip.flip_img(transformed_img, flip_type)
operate['flip_type'] = flip_type
return transformed_img, certain_transforms