def run():
start = time.time()
# Convert string of image data to uint8
if 'data' not in request.files:
return jsonify({'error': 'missing file param `data`'}), 400
data = request.files['data'].read()
if len(data) == 0:
return jsonify({'error': 'empty image'}), 400
# Convert string data to PIL Image
img = Image.open(io.BytesIO(data))
# Ensure i,qge size is under 1024
if img.size[0] > 1024 or img.size[1] > 1024:
img.thumbnail((1024, 1024))
# Process Image
res = u2net.run(np.array(img))
res = res.resize((img.size), resample=Image.BILINEAR) # remove resample
# empty_img = Image.new("RGBA", (img.size), 0)
bg = Image.open('bg.jpg')
bg = bg.resize((img.size), resample=Image.BILINEAR)
# new_img = Image.composite(img, empty_img, res.convert("L"))
new_img = Image.composite(img, bg, res.convert("L"))
# Save to buffer
buffer = io.BytesIO()
new_img.save(buffer, "PNG")
buffer.seek(0)
# Print stats
logging.info(f'Completed in {time.time() - start:.2f}s')
new_img.save("test.png")
return send_file(buffer, mimetype='image/png')
def run():
start = time.time()
# Convert string of image data to uint8
if 'data' not in request.files:
return jsonify({'error': 'missing file param `data`'}), 400
data = request.files['data'].read()
if len(data) == 0:
return jsonify({'error': 'empty image'}), 400
# Convert string data to PIL Image
img = Image.open(io.BytesIO(data))
# Ensure i,qge size is under 1024
if img.size[0] > 1024 or img.size[1] > 1024:
img.thumbnail((1024, 1024))
# Process Image
res = u2net.run(np.array(img))
# Save to buffer
buff = io.BytesIO()
res.save(buff, 'PNG')
buff.seek(0)
# Print stats
logging.info(f'Completed in {time.time() - start:.2f}s')
# Return data
return send_file(buff, mimetype='image/png')
def run():
# Convert string data to PIL Image
img = Image.open(io.BytesIO(b64decode(request.form['image'])))
# Process Image
res = u2net.run(img)
# Save to buffer
buff = io.BytesIO()
res.save(buff, 'PNG')
buff.seek(0)
# Return data
return send_file(buff, mimetype='image/png')
def run():
start = time.time()
# Convert string of image data to uint8
if 'data' not in request.files:
return jsonify({'error': 'missing file param `data`'}), 400
data = request.files['data'].read()
if len(data) == 0:
return jsonify({'error': 'empty image'}), 400
# Convert string data to PIL Image
img = Image.open(io.BytesIO(data))
# Ensure i,qge size is under 1024
if img.size[0] > 1024 or img.size[1] > 1024:
img.thumbnail((1024, 1024))
# Process Image
res = u2net.run(np.array(img))
# Convert to BW and make it to the size of original image
mask = res.convert('L').resize((img.width, img.height))
# Making final composite
logging.info(' > compositing final image...')
empty = Image.new("RGBA", img.size, 0)
img = Image.composite(img, empty, mask)
# Save to buffer
buff = io.BytesIO()
img.save(buff, 'PNG')
buff.seek(0)
# Print stats
logging.info(f'Completed in {time.time() - start:.2f}s')
# Return data
return send_file(buff, mimetype='image/png')
def process_frames(original_filename, output_filename, mask_filename,
thumbnail_filename):
# change to gpu(0) for faster processing
vr = VideoReader(original_filename, ctx=cpu(0))
height, width, layers = vr[0].shape
print(f'\u001b[33mInput frame {height}x{width}x{layers}\u001b[0m')
fourcc = cv2.VideoWriter_fourcc(*'FFV1')
video = cv2.VideoWriter(output_filename + '.lossless.mkv', fourcc,
vr.get_avg_fps(), (width, height))
video_mask = cv2.VideoWriter(mask_filename + '.lossless.mkv', fourcc,
vr.get_avg_fps(), (320, 320))
# solid color image
keying_bg = create_keying_background(width, height, (0, 255, 0))
for frame in vr:
# convert to numpy format
frame_np = frame.asnumpy()
# run u2net
mask_np = u2net.run(frame_np)
# write frame to mask video
mask_np_uint8 = (mask_np * 255).astype(np.uint8)
mask_np_bgr = np.stack([mask_np_uint8] * 3,
axis=-1) # https://stackoverflow.com/a/40119878
video_mask.write(mask_np_bgr)
# resize u2net output (320x320) to original frame resolution
mask_cv2 = cv2.resize(mask_np, (width, height))
# scale mask values from the range [0, 1] to [0, 255]
mask_cv2_uint8 = (mask_cv2 * 255).astype(np.uint8)
# thresholding the mask to have clear outlines
ret, mask_cv2_uint8 = cv2.threshold(mask_cv2_uint8, 10, 255,
cv2.THRESH_BINARY)
# compute inverse mask
mask_cv2_uint8_inv = cv2.bitwise_not(mask_cv2_uint8)
# apply mask to image and merge with keying background
frame_fg = cv2.bitwise_and(frame_np, frame_np, mask=mask_cv2_uint8)
frame_bg = cv2.bitwise_and(keying_bg,
keying_bg,
mask=mask_cv2_uint8_inv)
output_cv2 = frame_fg + frame_bg
# convert the color space back to BGR
output_cv2 = cv2.cvtColor(output_cv2, cv2.COLOR_RGB2BGR)
video.write(output_cv2)
cv2.destroyAllWindows()
video.release()
video_mask.release()
# encode videos to h264
thumbnail_proc = start_thumbnail(output_filename + '.lossless.mkv',
thumbnail_filename)
video_enc_proc = start_encode_video(output_filename + '.lossless.mkv',
output_filename)
video_mask_enc_proc = start_encode_video(mask_filename + '.lossless.mkv',
mask_filename)
assert thumbnail_proc.wait() == 0, 'Thumbnail encoding failed'
assert video_enc_proc.wait() == 0, 'Video encoding failed'
assert video_mask_enc_proc.wait() == 0, 'Mask video encoding failed'