def main():
"""
Inference function to generate SR images.
"""
nn.load_parameters(args.model)
# Inference data loader
inference_data = inference_data_loader(args.input_dir_lr)
input_shape = [
1,
] + list(inference_data.inputs[0].shape)
output_shape = [1, input_shape[1] * 4, input_shape[2] * 4, 3]
oh = input_shape[1] - input_shape[1] // 8 * 8
ow = input_shape[2] - input_shape[2] // 8 * 8
# Build the computation graph
inputs_raw = nn.Variable(input_shape)
pre_inputs = nn.Variable(input_shape)
pre_gen = nn.Variable(output_shape)
pre_warp = nn.Variable(output_shape)
transposed_pre_warp = space_to_depth(pre_warp)
inputs_all = F.concatenate(inputs_raw, transposed_pre_warp)
with nn.parameter_scope("generator"):
gen_output = generator(inputs_all, 3, args.num_resblock)
outputs = (gen_output + 1) / 2
inputs_frames = F.concatenate(pre_inputs, inputs_raw)
with nn.parameter_scope("fnet"):
flow_lr = flow_estimator(inputs_frames)
flow_lr = F.pad(flow_lr, (0, 0, 0, oh, 0, ow, 0, 0), "reflect")
flow_hr = upscale_four(flow_lr * 4.0)
pre_gen_warp = warp_by_flow(pre_gen, flow_hr)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
max_iter = len(inference_data.inputs)
print('Frame evaluation starts!!')
pre_inputs.d, pre_gen.d, pre_warp.d = 0, 0, 0
for i in range(max_iter):
inputs_raw.d = np.array([inference_data.inputs[i]]).astype(np.float32)
if i != 0:
pre_gen_warp.forward()
pre_warp.data.copy_from(pre_gen_warp.data)
outputs.forward()
output_frame = outputs.d
if i >= 5:
name, _ = os.path.splitext(
os.path.basename(str(inference_data.paths_lr[i])))
filename = args.output_name + '_' + name
print('saving image %s' % filename)
out_path = os.path.join(args.output_dir,
"%s.%s" % (filename, args.output_ext))
save_img(out_path, output_frame[0])
else: # First 5 is a hard-coded symmetric frame padding, ignored but time added!
print("Warming up %d" % (5 - i))
pre_inputs.data.copy_from(inputs_raw.data)
pre_gen.data.copy_from(outputs.data)
async def post_defect_detection(img: UploadFile = File(...)):
"""
The function first detects all objects present in an image
& then predicts whether or not each of them is defective
**Parameters:**
1. `img`: image file uploaded to the API endpoint using the POST request
"""
img_name = "temp-img.png"
label_req = ["pump_impeller"]
with open(img_name, "wb") as buffer:
shutil.copyfileobj(img.file, buffer)
"""
It first stores the uploaded image temporarily in the local database
of the server runs it through our Object Detection pipeline.
"""
detections = object_detection_model.detect_objects(img_name)
for i in range(len(detections)):
if detections[i]["label"] in label_req:
b_box = detections[i]["bounding_box"]
pred = extract_region_and_predict(img_name, b_box)
detections[i]["defect_pred"] = pred["prediction"]
"""
The bounding boxes of the detected parts are returned as a dictionary.
For each detected part, the region of interest is selected using the
bounding box & a prediction is made as to whether or not it contains
any defect based on our Defect Detection Model.
"""
final_img = utils.draw_bounding_boxes(img_name, detections)
"""
These predictions are used to draw colored bounding boxes around the detected parts
"""
os.remove(img_name)
with tempfile.NamedTemporaryFile(mode="w+b", suffix=".png",
delete=False) as FOUT:
utils.save_img(final_img, filename=FOUT.name)
encoded_image_string = base64.b64encode(FOUT.read())
return {
"mime": "image/png",
"image": encoded_image_string,
}
"""
def img_style_transfer(filename, filename2):
img = get_img(filename)
filename_dir = '/Users/tanya/PycharmProjects/image6/static/images/'
content_filename = filename_dir + filename
content_image = load_image(content_filename, max_size=None)
style_filename = filename_dir + filename2
style_image = load_image(style_filename, max_size=300)
content_layer_ids = [4]
style_layer_ids = list(range(13))
res = style_transfer(content_image=content_image,
style_image=style_image,
content_layer_ids=content_layer_ids,
style_layer_ids=style_layer_ids,
weight_content=5,
weight_style=10.0,
weight_denoise=0.3,
num_iterations=60,
step_size=10.0)
# 滤波处理
# 返回处理后文件名称和原始参数
return save_img(res, filename)
def rotate(filename: str, sp: list, size: int, degree: int):
if degree == 90:
angle = Image.ROTATE_90
elif degree == 180:
angle = Image.ROTATE_180
else:
angle = Image.ROTATE_270
# 获取图像
im = get_img(filename)
width, height = im.size
sp[0] = int(sp[0])
sp[1] = int(sp[1])
# 区域起点大于图像的尺寸
if sp[0] > width or sp[1] > height:
return filename
# 设置窗口大小
size_x = width - sp[0] if sp[0] + size > width else size
size_y = height - sp[1] if sp[1] + size > height else size
size = size_x if size_x < size_x else size_y
# 创建矩形窗,左-上-右-下
box = (sp[0], sp[1], sp[0] + size, sp[1] + size)
# 裁剪窗口
region = im.crop(box)
region = region.transpose(angle)
im.paste(region, box)
return save_img(im, filename)
def mosaic(filename: str, sp: list, ep: list, granularity: int) -> tuple:
olddata = '#'.join(sp + ep) + '#' + str(granularity)
# 获取图像
im = get_img(filename)
width, height = im.size
# 创建Draw对象
draw = ImageDraw.ImageDraw(im=im)
sp[0] = int(sp[0])
sp[1] = int(sp[1])
# 区域起点大于图像的尺寸
if sp[0] > width or sp[1] > height or granularity < 1:
return filename, olddata
# 处理终点
ep[0] = width if int(ep[0]) > width else int(ep[0])
ep[1] = height if int(ep[1]) > height else int(ep[1])
# 正式打马赛克
for x in range(int(sp[0]), ep[0], granularity):
for y in range(int(sp[1]), ep[1], granularity):
# 选取小方块中心的像素点填充整个小方块
r, g, b = im.getpixel((x, y))
draw.rectangle([(x, y),
(x + granularity - 1, y + granularity - 1)],
fill=(r, g, b))
# 保存并返回文件名
return save_img(im, filename), olddata
def img_filter(form, filename, **kwargs):
mode = kwargs.get('mode', None)
img = get_img(filename)
# 读取滤波方式
im_filter, olddata = get_filter(mode, form)
# 滤波处理
res = img.filter(im_filter)
# 返回处理后文件名称和原始参数
return save_img(res, filename), olddata
def img_kernel(filename, size, kernel, scale=None, offset=None):
im = get_img(filename)
if scale is not None and scale.strip() != '':
scale = float(scale)
else:
scale = None
if offset is not None and offset.strip() != '':
offset = float(offset)
else:
offset = 0
im_filter = ImageFilter.Kernel(size, kernel, scale, offset)
res = im.filter(im_filter)
return save_img(res, filename)
def img_enhance(filename, enhance_type, factor):
im = get_img(filename)
if enhance_type == 'Color':
enhancer = ImageEnhance.Color(im)
elif enhance_type == 'Contrast':
enhancer = ImageEnhance.Contrast(im)
elif enhance_type == 'Brightness':
enhancer = ImageEnhance.Brightness(im)
elif enhance_type == 'Sharpness':
enhancer = ImageEnhance.Sharpness(im)
else:
raise NotExistEnhanceError()
res = enhancer.enhance(factor)
return save_img(res, filename)
def overlay(filename1, filename2, alpha: float):
im1 = get_img(filename1)
im2 = get_img(filename2)
w = im1.size[0] if im1.size[0] > im2.size[0] else im2.size[0]
h = im1.size[1] if im1.size[1] > im2.size[1] else im2.size[1]
size = (w, h)
base1 = Image.new('RGB', size, ImageColor.getrgb('white'))
base2 = Image.new('RGB', size, ImageColor.getrgb('white'))
base1.paste(im1, (0, 0))
base2.paste(im2, (0, 0))
base1.show()
base2.show()
res = Image.blend(base1, base2, alpha)
res.show()
return save_img(res, filename1)
def text_mask(filename, text, color, font_size, start_point):
olddata = text + '#' + color + '#' + font_size + '#' + start_point[
0] + '#' + start_point[1]
im = get_img(filename)
draw = ImageDraw.ImageDraw(im)
font = ImageFont.truetype(font=basedir +
'/static/font/STHeiti-Light-3.ttc',
size=int(font_size),
encoding='utf-8')
if color in ImageColor.colormap:
color = ImageColor.getrgb(color)
else:
color = ImageColor.getrgb('red')
draw.text((int(start_point[0]), int(start_point[1])),
text,
fill=color,
font=font)
return save_img(im, filename), olddata
def water_mask(filename: str, mask_filename: str):
# 获取图像
im = get_img(filename)
width, height = im.size
# 获取水印图片
mask_filename = convert_img(filename, mask_filename)
mask_im = get_img(mask_filename)
mw, mh = mask_im.size
mw, mh = 250, int(mh * 250 / mw)
mask_im = mask_im.resize((mw, mh))
# 裁剪原来的图像区域并混叠
region = im.crop((width - mw, height - mh, width, height))
new = Image.blend(mask_im, region, 0.5)
im.paste(new, (width - mw, height - mh))
return save_img(im, filename)
print(log_str)
model.seen += imgs.size(0)
if epoch % 5 == 0:
predictions = non_max_suppression(outputs, conf_thres=opt.conf_thres, nms_thres=opt.nms_thres)
targets[:, 1:] = xywh2xyxy(targets[:, 1:])
# Rescale to image dimension
targets[:, 1:] *= opt.img_size
# Get batch statistics used to compute metrics
statistics = get_batch_statistics(predictions, targets, iou_threshold=0.5)
true_positives, pred_scores, pred_labels = [np.concatenate(x, 0) for x in list(zip(*statistics))]
# Compute metrics
precision, recall, AP, f1, ap_class = ap_per_class(
true_positives, pred_scores, pred_labels, list(range(int(data_config["classes"])))
)
global_metrics = [
("F1", f1.mean()),
("Recall", recall.mean()),
("Precision", precision.mean()),
("mAP", AP.mean()),
]
print(global_metrics)
for img_i, (img_path, detections) in enumerate(zip(paths, predictions)):
save_img(img_path, detections, classes, "output", str(epoch)+"_"+str(img_i), opt)
if epoch % opt.checkpoint_interval == 0 or epoch == opt.epochs-1:
model.save_weights("%s/%d.weights" % (opt.checkpoint_dir, epoch))
def save_img(self, fig, name):
utils.save_img(fig, self.model_name, name, self.result_dir)
return
def convert_img(dst, src):
dst_suffix = dst.split('.')[-1]
new_filename = src.split('.')[0] + '.' + dst_suffix
im = get_img(src)
im = im.convert('RGB')
return save_img(im, new_filename)
