Add production codes

app.py

@@ -0,0 +1,103 @@
+from flask import Flask, request, jsonify, render_template, make_response
+from flask_cors import CORS
+import numpy as np
+import cv2
+
+from src.color_controls import control_kelvin, control_contrast, control_HSV
+from src.cyano import Cyanotype
+from src.prediction import predict_img, optimize_img, update_patch
+from src.utils import cv_to_pil, pil_to_cv
+
+UPLOAD_FOLDER = './uploads'
+app = Flask(__name__, template_folder='/client', static_folder='/client')
+
+app.config['UPLOAD_FOLDER'] = UPLOAD_FOLDER
+
+CORS(
+    app,
+    supports_credentials=True
+)
+
+
+@app.route('/api/process', methods=['POST'])
+def process():
+    imgfile = request.files['img']
+    img_array = np.asarray(bytearray(imgfile.stream.read()), dtype=np.uint8)
+    img = cv2.imdecode(img_array, cv2.IMREAD_COLOR)
+
+    data = request.form
+    hue = int(data["hue"])
+    saturation = float(data["saturation"])
+    lightness = float(data["lightness"])
+    contrast = int(data["contrast"])
+    kelvin = int(data["kelvin"])
+
+    img = control_contrast(img, contrast)
+    img = control_HSV(img, hue, saturation, lightness)
+
+    img_pil = cv_to_pil(img)
+    img_pil = control_kelvin(img_pil, kelvin)
+    img = pil_to_cv(img_pil)
+
+    response = make_response(cv2.imencode('.png', img)[1].tobytes())
+    response.headers.set('Content-Type', 'image/png')
+
+    return response
+
+
+@app.route('/api/predict/<process_name>', methods=['POST'])
+def predict(process_name):
+    if not process_name in ['cyanotype_mono', 'cyanotype_full', 'salt', 'platinum']:
+        return jsonify({ 'error': 'process name is invalid' })
+
+    imgfile = request.files['img']
+    img_array = np.asarray(bytearray(imgfile.stream.read()), dtype=np.uint8)
+    img = cv2.imdecode(img_array, cv2.IMREAD_COLOR)
+
+    if 'colorpatch' in request.files:
+        patchfile = request.files['colorpatch']
+        patch_array = np.asarray(bytearray(patchfile.stream.read()), dtype=np.uint8)
+        colorpatch = cv2.imdecode(colorpatch_array, cv2.IMREAD_COLOR)
+        update_patch(process_name, colorpatch)
+
+    img = predict_img(process_name, img)
+
+    response = make_response(cv2.imencode('.png', img)[1].tobytes())
+    response.headers.set('Content-Type', 'image/png')
+
+    return response
+
+
+@app.route('/api/optimize/<process_name>', methods=['POST'])
+def optimize(process_name):
+    if not process_name in ['cyanotype_mono', 'cyanotype_full', 'salt', 'platinum']:
+        return jsonify({ 'error': 'process name is invalid' })
+
+    imgfile = request.files['img']
+    img_array = np.asarray(bytearray(imgfile.stream.read()), dtype=np.uint8)
+    img = cv2.imdecode(img_array, cv2.IMREAD_COLOR)
+
+    if 'colorpatch' in request.files:
+        patchfile = request.files['colorpatch']
+        patch_array = np.asarray(bytearray(patchfile.stream.read()), dtype=np.uint8)
+        colorpatch = cv2.imdecode(colorpatch_array, cv2.IMREAD_COLOR)
+        update_patch(process_name, colorpatch)
+
+    (opt_img, preview_img) = optimize_img(process_name, img)
+
+    h, w = preview_img.shape[:2]
+    if process_name.endswith('full'):
+        opt_img = np.reshape(opt_img, (h, w, 3))
+    else:
+        opt_img = np.reshape(opt_img, (h, w, 1))
+        opt_img = np.array([[[i[0]] * 3 for i in j] for j in opt_img], dtype=np.uint8)
+
+    img = cv2.hconcat([opt_img, preview_img])
+    response = make_response(cv2.imencode('.png', img)[1].tobytes())
+    response.headers.set('Content-Type', 'image/png')
+
+    return response
+
+
+if __name__ == "__main__":
+    app.run(debug=True, port=8000)

requirements.txt

@@ -0,0 +1,59 @@
+-i https://pypi.org/simple
+absl-py==1.2.0; python_version >= '3.6'
+astunparse==1.6.3
+cachetools==5.2.0; python_version ~= '3.7'
+certifi==2022.6.15; python_version >= '3.6'
+charset-normalizer==2.1.0; python_version >= '3.6'
+click==8.1.3; python_version >= '3.7'
+flask==2.2.2
+flask-cors==3.0.10
+flatbuffers==1.12
+gast==0.4.0; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'
+google-auth==2.10.0; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4, 3.5'
+google-auth-oauthlib==0.4.6; python_version >= '3.6'
+google-pasta==0.2.0
+grpcio==1.47.0; python_version >= '3.6'
+gunicorn==20.1.0
+h5py==3.7.0; python_version >= '3.7'
+idna==3.3; python_version >= '3.5'
+importlib-metadata==4.12.0; python_version < '3.10'
+itsdangerous==2.1.2; python_version >= '3.7'
+jinja2==3.1.2; python_version >= '3.7'
+joblib==1.1.0; python_version >= '3.6'
+keras==2.9.0
+keras-preprocessing==1.1.2
+libclang==14.0.6
+markdown==3.4.1; python_version >= '3.7'
+markupsafe==2.1.1; python_version >= '3.7'
+numpy==1.21.6
+oauthlib==3.2.0; python_version >= '3.6'
+opencv-python==4.6.0.66
+opt-einsum==3.3.0; python_version >= '3.5'
+packaging==21.3; python_version >= '3.6'
+pandas==1.4.2
+pillow==9.2.0
+protobuf==3.19.4; python_version >= '3.5'
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pyparsing==3.0.9; python_full_version >= '3.6.8'
+requests==2.28.1; python_version >= '3.7' and python_version < '4'
+requests-oauthlib==1.3.1; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'
+rsa==4.9; python_version >= '3.6'
+scikit-learn==1.0.2
+scipy==1.7.3; python_version < '3.11' and python_version >= '3.7'
+setuptools==65.0.0; python_version >= '3.7'
+six==1.16.0; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'
+tensorboard==2.9.1; python_version >= '3.6'
+tensorboard-data-server==0.6.1; python_version >= '3.6'
+tensorboard-plugin-wit==1.8.1
+tensorflow==2.9.1
+tensorflow-estimator==2.9.0; python_version >= '3.7'
+tensorflow-io-gcs-filesystem==0.26.0; python_version < '3.11' and python_version >= '3.7'
+termcolor==1.1.0
+threadpoolctl==3.1.0; python_version >= '3.6'
+typing-extensions==4.3.0; python_version >= '3.7'
+urllib3==1.26.11; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4, 3.5' and python_version < '4'
+werkzeug==2.2.2; python_version >= '3.7'
+wheel==0.37.1; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4'
+wrapt==1.14.1; python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4'
+zipp==3.8.1; python_version >= '3.7'

src/color_controls.py

@@ -0,0 +1,89 @@
+import cv2
+import numpy as np
+import math
+
+# HSV（色相，彩度，明度）を制御する関数
+def control_HSV(img, h_deg, s_mag, v_mag):
+    img_hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
+    print(h_deg, s_mag, v_mag)
+    # HSVの回転
+    img_hsv[:,:,(0)] = img_hsv[:,:,(0)] + h_deg
+    img_hsv[:,:,(1)] = img_hsv[:,:,(1)] * s_mag
+    img_hsv[:,:,(2)] = img_hsv[:,:,(2)] * v_mag
+    # HSV to RGB
+    img_bgr = cv2.cvtColor(img_hsv, cv2.COLOR_HSV2BGR)
+    return img_bgr
+
+# コントラストを制御する関数
+def control_contrast(img, contrast):
+    #コントラスト調整ファクター
+    factor = (259 *(contrast + 255)) / (255 *(259 - contrast))
+    #float型に変換
+    newImage = np.array(img, dtype = 'float64')
+    #コントラスト調整。（0以下 or 255以上）はクリッピング
+    newImage = np.clip((newImage[:,:,:] - 128) * factor + 128, 0, 255)
+    #int型に戻す
+    newImage = np.array(newImage, dtype = 'uint8')
+
+    return newImage
+
+
+def __clamp(value: float, min_val: int = 0, max_val: int = 255) -> int:
+    # use rounding to better represent values between max and min
+    return int(round(max(min(value, max_val), min_val)))
+
+
+def kelvin_to_rgb(kelvin):
+    temperature = kelvin / 100.0
+
+    if temperature < 66.0:
+        red = 255
+    else:
+        # a + b x + c Log[x] /.
+        # {a -> 351.97690566805693`,
+        # b -> 0.114206453784165`,
+        # c -> -40.25366309332127
+        # x -> (kelvin/100) - 55}
+        red = temperature - 55.0
+        red = 351.97690566805693 + 0.114206453784165 * red - 40.25366309332127 * math.log(red)
+
+    # Calculate green
+    if temperature < 66.0:
+        # a + b x + c Log[x] /.
+        # {a -> -155.25485562709179`,
+        # b -> -0.44596950469579133`,
+        # c -> 104.49216199393888`,
+        # x -> (kelvin/100) - 2}
+        green = temperature - 2
+        green = -155.25485562709179 - 0.44596950469579133 * green + 104.49216199393888 * math.log(green)
+    else:
+        # a + b x + c Log[x] /.
+        # {a -> 325.4494125711974`,
+        # b -> 0.07943456536662342`,
+        # c -> -28.0852963507957`,
+        # x -> (kelvin/100) - 50}
+        green = temperature - 50.0
+        green = 325.4494125711974 + 0.07943456536662342 * green - 28.0852963507957 * math.log(green)
+
+    # Calculate blue
+    if temperature >= 66.0:
+        blue = 255
+    elif temperature <= 20.0:
+        blue = 0
+    else:
+        # a + b x + c Log[x] /.
+        # {a -> -254.76935184120902`,
+        # b -> 0.8274096064007395`,
+        # c -> 115.67994401066147`,
+        # x -> kelvin/100 - 10}
+        blue = temperature - 10
+        blue = -254.76935184120902 + 0.8274096064007395 * blue + 115.67994401066147 * math.log(blue)
+
+    return __clamp(red, 0, 255), __clamp(blue, 0, 255), __clamp(green, 0, 255)
+
+def control_kelvin(img, kelvin):
+    r, g, b = kelvin_to_rgb(int(kelvin))
+    color_matrix = (r / 255.0, 0.0, 0.0, 0.0, 0.0, g / 255.0, 0.0, 0.0, 0.0, 0.0, b / 255.0, 0.0)
+    new_img = img.convert('RGB', color_matrix)
+
+    return new_img

src/cyano.py

@@ -0,0 +1,151 @@
+import cv2
+import numpy as np
+import tensorflow as tf
+from sklearn.linear_model import LinearRegression
+
+from src.get_patch_rgb import get_patch_rgb
+
+
+class Cyanotype():
+    def __init__(self):
+        ### set for patch ###
+        patch_img_path  = './colorpatches/cyanotype_full.png'
+        self.update_patch(cv2.imread(patch_img_path, cv2.IMREAD_COLOR))
+
+
+    def update_patch(self, patch_img):
+        self.rgb_cyano  = [[0,0,0]]
+        self.patch_img = patch_img
+        self.patch_img_height, self.patch_img_width, _ = self.patch_img.shape
+
+        self.crop_img()
+
+        self.patch_rgb  = get_patch_rgb()
+        self.patch_rgb  = np.array(self.patch_rgb)
+        # self.patch_rgb    = self.patch_rgb/255.0
+        self.rgb_cyano  = np.array(self.rgb_cyano)
+        # self.rgb_cyano    = self.rgb_cyano/255.0
+        print(self.patch_rgb.shape)
+        print(self.rgb_cyano.shape)
+
+        self.fit_model()
+
+
+    def crop_img(self):
+        # 対象範囲を切り出し
+        h_pix = 14
+        w_pix = 21
+        w_ = round(self.patch_img_width/w_pix)
+        h_ = round(self.patch_img_height/h_pix)
+        for i in range(h_pix):
+            for j in range(w_pix):
+                boxFromX = j*w_+5 #対象範囲開始位置 X座標
+                boxFromY = i*h_+5 #対象範囲開始位置 Y座標
+                boxToX = ((j+1)*w_)-7 #対象範囲終了位置 X座標
+                boxToY = ((i+1)*h_)-7 #対象範囲終了位置 Y座標
+                # y:y+h, x:x+w の順で設定
+                imgBox = self.patch_img[boxFromY: boxToY, boxFromX: boxToX]
+
+                # RGB平均値を出力
+                # flattenで一次元化しmeanで平均を取得
+                b = imgBox.T[0].flatten().mean()
+                g = imgBox.T[1].flatten().mean()
+                r = imgBox.T[2].flatten().mean()
+
+                self.rgb_cyano.append([r,g,b])
+
+        del self.rgb_cyano[0]
+
+
+    def fit_model(self):
+        self.reg = LinearRegression().fit(self.patch_rgb, self.rgb_cyano)
+        self.reg.score(self.patch_rgb, self.rgb_cyano)
+        print('self.reg.coef_: ', self.reg.coef_)
+        print('self.reg.intercept_: ', self.reg.intercept_)
+
+
+    def predict_img(self, img):
+        print(img.shape)
+        img         = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        # img         = cv2.resize(img, dsize=None, fx=0.2, fy=0.2)
+        # img         = cv2.resize(img, dsize=(100, 100))
+        # img         = cv2.resize(img, dsize=(500, 500))
+
+        img_cyano   = img @ self.reg.coef_.T + self.reg.intercept_
+        img_cyano   = img_cyano.astype(np.uint8)
+        img_cyano   = cv2.cvtColor(img_cyano, cv2.COLOR_RGB2BGR)
+        img_cyano   = np.array(img_cyano)
+        print(img_cyano.shape)
+
+        return img_cyano
+
+
+    def MSE(self, imageA, imageB):
+        err = np.sum((imageA.astype("float") - imageB.astype("float")) ** 2)
+        err /= float(imageA.shape[0] * imageA.shape[1] * imageA.shape[2])
+        return err
+
+
+    # ---------- Optimization with Tensorflow ---------- #
+    def tf_optimize(self, img):
+        print('\n---------- Start Optimization ----------')
+        img         = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        x           = self.reg.coef_
+        A           = img
+        target      = img
+        A_height    = A.shape[0]
+        A_width     = A.shape[1]
+        cnt         = A_height*A_width
+        print(A.shape)
+        print(cnt)
+
+        param_tf      = tf.Variable(A, dtype=tf.float64)
+        coef_tf       = tf.constant(x.T, dtype=tf.float64)
+        intercept_tf  = tf.constant(self.reg.intercept_, dtype=tf.float64)
+        target_tf     = tf.constant(target, dtype=tf.float64)
+
+        opt = tf.keras.optimizers.Adam(learning_rate=5.0)
+#         opt = tf.keras.optimizers.Adam(learning_rate=0.1)
+
+        def loss():
+            x0        = param_tf
+            x0        = tf.where(x0 > 255.0, 255.0, x0)
+            x0        = tf.where(x0 < 0.0, 0.0, x0)
+            x0        = tf.reshape(x0, [cnt, 3])
+            t_tf      = target_tf
+            t_tf      = tf.reshape(t_tf, [cnt, 3])
+            pred      = tf.linalg.matmul(x0, coef_tf) + intercept_tf
+            diff      = pred - t_tf
+            diff_2    = diff**2
+            pix_cnt   = tf.size(t_tf)
+            pix_cnt   = tf.cast(pix_cnt, dtype=tf.float64)
+            loss_val  = tf.math.reduce_sum(diff_2) / pix_cnt
+            print('loss_val: ', loss_val)
+            return loss_val
+
+        for i in range(50):
+            step_count = opt.minimize(loss, [param_tf]).numpy()
+            # if step_count==10:
+            #   break
+            print(step_count)
+
+        # ----- check optimized result ----- #
+        x0 = param_tf
+        x0 = tf.where(x0 > 255.0, 255.0, x0)
+        x0 = tf.where(x0 < 0.0, 0.0, x0)
+        x0 = x0.numpy()
+        x0 = x0.reshape((cnt, 3))
+        sim_opt = x0 @ x.T + self.reg.intercept_
+        sim_opt = sim_opt.reshape((A_height, A_width, 3))
+        sim_opt = sim_opt.astype(np.uint8)
+        sim_opt = cv2.cvtColor(sim_opt, cv2.COLOR_RGB2BGR)
+
+        return (x0, sim_opt)
+
+
+if __name__ == '__main__':
+    img = cv2.imread('samples/input/00.jpg', cv2.IMREAD_COLOR)
+    cy = Cyanotype()
+    cy.fit_model()
+    cy.predict_img(img)
+    cy.tf_optimize(img)

src/get_patch_rgb.py

@@ -0,0 +1,3 @@
+def get_patch_rgb():
+    rgb = [[0.0000,0.0000,0.0000],[25.000,0.0000,0.0000],[51.000,0.0000,0.0000],[76.000,0.0000,0.0000],[102.00,0.0000,0.0000],[127.00,0.0000,0.0000],[153.00,0.0000,0.0000],[178.00,0.0000,0.0000],[204.00,0.0000,0.0000],[229.00,0.0000,0.0000],[255.00,0.0000,0.0000],[0.0000,25.000,0.0000],[0.0000,51.000,0.0000],[0.0000,76.000,0.0000],[0.0000,102.00,0.0000],[0.0000,127.00,0.0000],[0.0000,153.00,0.0000],[0.0000,178.00,0.0000],[0.0000,204.00,0.0000],[0.0000,229.00,0.0000],[0.0000,255.00,0.0000],[0.0000,0.0000,25.000],[0.0000,0.0000,51.000],[0.0000,0.0000,76.000],[0.0000,0.0000,102.00] ,[0.0000,0.0000,127.00],[0.0000,0.0000,153.00] ,[0.0000,0.0000,178.00],[0.0000,0.0000,204.00],[0.0000,0.0000,229.00],[0.0000,0.0000,255.00] ,[255.00,255.00,0.0000],[255.00,0.0000,255.00],[0.0000,255.00,255.00],[255.00,255.00,255.00],[255.00,255.00,191.00],[255.00,255.00,127.00],[255.00,255.00,63.000],[255.00,191.00,255.00],[255.00,191.00,191.00],[255.00,191.00,127.00] ,[255.00,191.00,63.000],[255.00,191.00,0.0000] ,[255.00,127.00,255.00],[255.00,127.00,191.00],[255.00,127.00,127.00] ,[255.00,127.00,63.000],[255.00,127.00,0.0000] ,[255.00,63.000,255.00],[255.00,63.000,191.00],[255.00,63.000,127.00],[255.00,63.000,63.000],[255.00,63.000,0.0000],[255.00,0.0000,191.00],[255.00,0.0000,127.00],[255.00,0.0000,63.000],[191.00,255.00,255.00],[191.00,255.00,191.00],[191.00,255.00,127.00],[191.00,255.00,63.000],[191.00,255.00,0.0000],[191.00,191.00,255.00],[191.00,191.00,191.00],[191.00,191.00,127.00],[191.00,191.00,63.000],[191.00,191.00,0.0000],[191.00,127.00,255.00],[191.00,127.00,191.00],[191.00,127.00,127.00],[191.00,127.00,63.000],[191.00,127.00,0.0000],[191.00,63.000,255.00],[191.00,63.000,191.00],[191.00,63.000,127.00],[191.00,63.000,63.000],[191.00,63.000,0.0000],[191.00,0.0000,255.00] ,[191.00,0.0000,191.00] ,[191.00,0.0000,127.00] ,[191.00,0.0000,63.000] ,[191.00,0.0000,0.0000] ,[127.00,255.00,255.00] ,[127.00,255.00,191.00] ,[127.00,255.00,127.00] ,[127.00,255.00,63.000],[127.00,255.00,0.0000] ,[127.00,191.00,255.00] ,[127.00,191.00,191.00] ,[127.00,191.00,127.00] ,[127.00,191.00,63.000] ,[127.00,191.00,0.0000],[127.00,127.00,255.00],[127.00,127.00,191.00] ,[127.00,127.00,127.00] ,[127.00,127.00,63.000] ,[127.00,127.00,0.0000] ,[127.00,63.000,255.00] ,[127.00,63.000,191.00],[127.00,63.000,127.00],[127.00,63.000,63.000] ,[127.00,63.000,0.0000] ,[127.00,0.0000,255.00],[127.00,0.0000,191.00] ,[127.00,0.0000,127.00] ,[127.00,0.0000,63.000] ,[63.000,255.00,255.00],[63.000,255.00,191.00],[63.000,255.00,127.00] ,[63.000,255.00,63.000] ,[63.000,255.00,0.0000] ,[63.000,191.00,255.00],[63.000,191.00,191.00] ,[63.000,191.00,127.00] ,[63.000,191.00,63.000],[63.000,191.00,0.0000],[63.000,127.00,255.00] ,[63.000,127.00,191.00] ,[63.000,127.00,127.00] ,[63.000,127.00,63.000] ,[63.000,127.00,0.0000] ,[63.000,63.000,255.00] ,[63.000,63.000,191.00] ,[63.000,63.000,127.00] ,[63.000,63.000,63.000] ,[63.000,63.000,0.0000] ,[63.000,0.0000,255.00] ,[63.000,0.0000,191.00] ,[63.000,0.0000,127.00] ,[63.000,0.0000,63.000] ,[63.000,0.0000,0.0000] ,[0.0000,255.00,191.00],[0.0000,255.00,127.00] ,[0.0000,255.00,63.000] ,[0.0000,191.00,255.00] ,[0.0000,191.00,191.00] ,[0.0000,191.00,127.00] ,[0.0000,191.00,63.000] ,[0.0000,191.00,0.0000] ,[0.0000,127.00,255.00],[0.0000,127.00,191.00],[0.0000,127.00,127.00] ,[0.0000,127.00,63.000],[0.0000,63.000,255.00] ,[0.0000,63.000,191.00] ,[0.0000,63.000,127.00] ,[0.0000,63.000,63.000] ,[0.0000,63.000,0.0000],[0.0000,0.0000,191.00],[0.0000,0.0000,63.000] ,[234.00,255.00,255.00] ,[211.00,255.00,255.00] ,[158.00,255.00,255.00] ,[255.00,234.00,255.00],[255.00,211.00,255.00] ,[255.00,158.00,255.00] ,[255.00,255.00,234.00],[255.00,255.00,211.00],[255.00,255.00,158.00],[20.000,255.00,255.00],[43.000,255.00,255.00],[96.000,255.00,255.00],[0.0000,234.00,255.00],[0.0000,211.00,255.00],[0.0000,158.00,255.00],[0.0000,255.00,234.00],[0.0000,255.00,211.00],[0.0000,255.00,158.00],[234.00,0.0000,255.00],[211.00,0.0000,255.00],[158.00,0.0000,255.00],[255.00,20.000,255.00],[255.00,43.000,255.00],[255.00,96.000,255.00],[255.00,0.0000,234.00],[255.00,0.0000,211.00],[255.00,0.0000,158.00],[20.000,0.0000,255.00],[43.000,0.0000,255.00],[96.000,0.0000,255.00],[0.0000,20.000,255.00],[0.0000,43.000,255.00],[0.0000,96.000,255.00],[0.0000,0.0000,234.00],[0.0000,0.0000,211.00],[0.0000,0.0000,158.00],[234.00,255.00,0.0000],[211.00,255.00,0.0000],[158.00,255.00,0.0000],[255.00,234.00,0.0000],[255.00,211.00,0.0000],[255.00,158.00,0.0000],[255.00,255.00,20.000],[255.00,255.00,43.000],[255.00,255.00,96.000],[20.000,255.00,0.0000],[43.000,255.00,0.0000],[96.000,255.00,0.0000],[0.0000,234.00,0.0000],[0.0000,211.00,0.0000],[0.0000,158.00,0.0000],[0.0000,255.00,20.000],[0.0000,255.00,43.000],[0.0000,255.00,96.000],[234.00,0.0000,0.0000],[211.00,0.0000,0.0000],[158.00,0.0000,0.0000],[255.00,20.000,0.0000],[255.00,43.000,0.0000],[255.00,96.000,0.0000],[255.00,0.0000,20.000],[255.00,0.0000,43.000],[255.00,0.0000,96.000],[20.000,0.0000,0.0000],[43.000,0.0000,0.0000],[96.000,0.0000,0.0000],[0.0000,20.000,0.0000],[0.0000,43.000,0.0000],[0.0000,96.000,0.0000],[0.0000,0.0000,20.000],[0.0000,0.0000,43.000],[0.0000,0.0000,96.000],[30.000,30.000,30.000],[94.000,94.000,94.000],[158.00,158.00,158.00],[221.00,221.00,221.00],[15.000,0.0000,0.0000],[0.0000,15.000,0.0000],[0.0000,0.0000,15.000],[0.0000,15.000,15.000],[15.000,0.0000,15.000],[15.000,15.000,0.0000],[48.000,30.000,30.000],[30.000,48.000,30.000],[30.000,30.000,48.000],[30.000,48.000,48.000],[48.000,30.000,48.000],[48.000,48.000,30.000],[79.000,63.000,63.000],[63.000,79.000,63.000],[63.000,63.000,79.000],[63.000,79.000,79.000],[79.000,63.000,79.000],[79.000,79.000,63.000],[112.00,94.000,94.000],[94.000,112.00,94.000],[94.000,94.000,112.00],[94.000,112.00,112.00],[112.00,94.000,112.00],[112.00,112.00,94.000],[142.00,127.00,127.00],[127.00,142.00,127.00],[127.00,127.00,142.00],[127.00,142.00,142.00],[142.00,127.00,142.00],[142.00,142.00,127.00],[175.00,158.00,158.00],[158.00,175.00,158.00],[158.00,158.00,175.00],[158.00,175.00,175.00],[175.00,158.00,175.00],[175.00,175.00,158.00],[206.00,191.00,191.00],[191.00,206.00,191.00],[191.00,191.00,206.00],[191.00,206.00,206.00],[206.00,191.00,206.00],[206.00,206.00,191.00],[239.00,221.00,221.00],[221.00,239.00,221.00],[221.00,221.00,239.00],[221.00,239.00,239.00],[239.00,221.00,239.00],[239.00,239.00,221.00],[30.000,0.0000,0.0000],[0.0000,30.000,0.0000],[0.0000,0.0000,30.000],[0.0000,30.000,30.000],[30.000,0.0000,30.000],[30.000,30.000,0.0000],[94.000,63.000,63.000],[63.000,94.000,63.000],[63.000,63.000,94.000],[63.000,94.000,94.000],[255.00,255.00,255.00],[255.00,255.00,255.00],[255.00,255.00,255.00],[255.00,255.00,255.00],[255.00,255.00,255.00],[255.00,255.00,255.00],[255.00,255.00,255.00],[255.00,255.00,255.00],[255.00,255.00,255.00],[255.00,255.00,255.00],[255.00,255.00,255.00]]
+    return rgb

src/mono_alternative.py

@@ -0,0 +1,213 @@
+import os
+import cv2
+import numpy as np
+import tensorflow as tf
+import pandas as pd
+from sklearn.linear_model import LinearRegression
+
+
+class MonoAlternative():
+    def __init__(
+        self,
+        process_name,
+        debug = False,
+    ):
+        patch_img_path = f'./colorpatches/{process_name}.png'
+        self.debug = debug
+        self.update_patch(cv2.imread(patch_img_path, cv2.IMREAD_COLOR))
+
+
+    def update_patch(self, patch_img):
+        self.patch_img  = patch_img
+        self.patch_img  = cv2.resize(self.patch_img, (512,512))
+        self.patch_img_height, self.patch_img_width, _ = self.patch_img.shape
+
+        self.cyano_rgb  = [[0,0,0]]
+        self.crop_img()
+        self.cyano_rgb  = np.array(self.cyano_rgb)
+
+        self.patch_rgb  = self.create_patch_arr()
+        self.patch_rgb  = np.array(self.patch_rgb)
+        print(self.cyano_rgb.shape)
+        print(self.patch_rgb.shape)
+
+        self.create_LUT()
+        self.fit_model()
+
+
+    def create_patch_arr(self):
+        patch_arr = np.empty((256, 3))
+        for i in range(256):
+            patch_arr[i] = np.array([i, i, i])
+        return patch_arr
+
+
+    def save_cropped_img(self, img, cnt):
+        patch_dir = './patch_data/'
+        if not os.path.exists(patch_dir):
+            os.makedirs(patch_dir)
+        cv2.imwrite(patch_dir + "patch_" + str(cnt) + ".png", img)
+
+
+    def create_LUT(self):
+        self.lut_arr = np.hstack([self.patch_rgb, self.cyano_rgb])
+        print('self.lut_arr.shape: ', self.lut_arr.shape)
+        df = pd.DataFrame(self.lut_arr, columns=['r','g','b','r_','g_','b_'])
+        print(df)
+        df.to_csv('./lut.csv')
+
+
+    def crop_img(self):
+        # 対象範囲を切り出し
+        h_pix = 16
+        w_pix = 16
+        w_ = round(self.patch_img_width/w_pix)
+        h_ = round(self.patch_img_height/h_pix)
+        for i in range(w_pix):
+            for j in range(h_pix):
+                boxFromX = i*w_+5 #対象範囲開始位置 X座標
+                boxFromY = j*h_+5 #対象範囲開始位置 Y座標
+                boxToX = ((i+1)*w_)-12 #対象範囲終了位置 X座標
+                boxToY = ((j+1)*h_)-12 #対象範囲終了位置 Y座標
+                # y:y+h, x:x+w の順で設定
+                imgBox = self.patch_img[boxFromY: boxToY, boxFromX: boxToX]
+                if self.debug:
+                    cnt = i*h_pix+j
+                    self.save_cropped_img(imgBox, cnt)
+
+                # RGB平均値を出力
+                # flattenで一次元化しmeanで平均を取得
+                b = imgBox.T[0].flatten().mean()
+                g = imgBox.T[1].flatten().mean()
+                r = imgBox.T[2].flatten().mean()
+
+                self.cyano_rgb.append([r,g,b])
+
+        del self.cyano_rgb[0]
+
+
+    def predict_img_LUT(self, img):
+        '''
+        img: color img (have to convert it to grayscale one)
+        return> bgr img with predicted cyano color
+        '''
+        print('img.shape:', img.shape)
+
+        if len(img.shape)==3: # if img is color
+            img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+        else:
+            img_gray = img
+        
+        print('img_gray.shape:', img.shape)
+        h, w        = img_gray.shape
+        pred_img    = np.empty((h, w, 3))
+        for i in range(h):
+            for j in range(w):
+                pix_val = img_gray[i][j]
+                pred_img[i][j] = self.lut_arr[pix_val, 3:6]
+
+        print('pred_img.shape:', pred_img.shape)
+        pred_img_bgr = cv2.cvtColor(pred_img.astype(np.float32), cv2.COLOR_RGB2BGR)
+
+        return pred_img_bgr, img_gray
+
+
+    def MSE(self, imageA, imageB):
+        err = np.sum((imageA.astype("float") - imageB.astype("float")) ** 2)
+        err /= float(imageA.shape[0] * imageA.shape[1] * imageA.shape[2])
+        return err
+
+
+    def fit_model(self):
+        self.patch_gray = np.array([self.patch_rgb[:, 0]]).reshape((256,1))
+        self.reg = LinearRegression().fit(self.patch_gray, self.cyano_rgb)
+        self.reg.score(self.patch_gray, self.cyano_rgb)
+        print('self.reg.coef_: ', self.reg.coef_)
+        print('self.reg.intercept_: ', self.reg.intercept_)
+
+
+    def predict_img(self, img):
+        print(img.shape)
+        h = img.shape[0]
+        w = img.shape[1]
+        if len(img.shape) == 3 and img.shape[2] == 3:
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+        img = img.reshape((h,w,1))
+        print(img.shape)
+        print(self.reg.coef_.T.shape)
+        img_cyano   = img @ self.reg.coef_.T + self.reg.intercept_
+        img_cyano   = img_cyano.astype(np.uint8)
+        img_cyano   = cv2.cvtColor(img_cyano, cv2.COLOR_RGB2BGR)
+        img_cyano   = np.array(img_cyano)
+        print(img_cyano.shape)
+
+        return img_cyano
+
+
+    # ---------- Optimization with Tensorflow ---------- #
+    def tf_optimize(self, img):
+        '''
+        img: 1 channel gray
+        but, target is 3 channel output
+        '''
+        print('\n---------- Start Optimization ----------')
+        img_3ch     = np.stack((img,)*3, axis=-1)
+        x           = self.reg.coef_
+        A           = img
+        target      = img_3ch
+        A_height    = A.shape[0]
+        A_width     = A.shape[1]
+        cnt         = A_height*A_width
+        print(A.shape)
+        print(cnt)
+
+        param_tf      = tf.Variable(A, dtype=tf.float64)
+        coef_tf       = tf.constant(x.T, dtype=tf.float64)
+        intercept_tf  = tf.constant(self.reg.intercept_, dtype=tf.float64)
+        target_tf     = tf.constant(target, dtype=tf.float64)
+
+        opt = tf.keras.optimizers.Adam(learning_rate=5.0)
+#         opt = tf.keras.optimizers.Adam(learning_rate=0.1)
+
+        def loss():
+            x0        = param_tf
+            x0        = tf.where(x0 > 255.0, 255.0, x0)
+            x0        = tf.where(x0 < 0.0, 0.0, x0)
+            x0        = tf.reshape(x0, [cnt, 1])
+            t_tf      = target_tf
+            t_tf      = tf.reshape(t_tf, [cnt, 3])
+            pred      = tf.linalg.matmul(x0, coef_tf) + intercept_tf
+            diff      = pred - t_tf
+            diff_2    = diff**2
+            pix_cnt   = tf.size(t_tf)
+            pix_cnt   = tf.cast(pix_cnt, dtype=tf.float64)
+            loss_val  = tf.math.reduce_sum(diff_2) / pix_cnt
+            print('loss_val: ', loss_val)
+            return loss_val
+
+        for i in range(50):
+            step_count = opt.minimize(loss, [param_tf]).numpy()
+            # if step_count==10:
+            #   break
+            print(step_count)
+
+        # ----- check optimized result ----- #
+        x0 = param_tf
+        x0 = tf.where(x0 > 255.0, 255.0, x0)
+        x0 = tf.where(x0 < 0.0, 0.0, x0)
+        x0 = x0.numpy()
+        x0_1d   = x0.reshape((cnt, 1))
+        sim_opt = x0_1d @ x.T + self.reg.intercept_
+        sim_opt = sim_opt.reshape((A_height, A_width, 3))
+        sim_opt = sim_opt.astype(np.uint8)
+        sim_opt = cv2.cvtColor(sim_opt, cv2.COLOR_RGB2BGR)
+
+        return (x0, sim_opt)
+
+
+
+if __name__ == '__main__':
+    cy = MonoAlternative()
+    cy.fit_model()
+    cy.predict_img()
+    cy.tf_optimize()

src/prediction.py

@@ -0,0 +1,68 @@
+import os
+import re
+from glob import glob
+import cv2
+from src.cyano import Cyanotype
+from src.mono_alternative import MonoAlternative
+
+print('Fitting models...')
+
+models = {
+  'cyanotype_full': Cyanotype(),
+  'cyanotype_mono': MonoAlternative('cyanotype_mono'),
+  'salt': MonoAlternative('salt'),
+  'platinum': MonoAlternative('platinum'),
+}
+
+
+def get_suffix_number(directory):
+  files = glob(f'{directory}/*.png')
+  suffixes = [re.search(r'[0-9]+', f) for f in files]
+  return max([int(s.group()) for s in suffixes if s] + [0]) + 1
+
+
+def update_patch(process_name, colorpatch):
+    model = models[process_name]
+    model.update_patch(colorpatch)
+    model.fit_model()
+
+
+def predict_img(process_name, img):
+  out_dir = f'outputs/{process_name}'
+  if not os.path.exists(out_dir):
+    os.makedirs(out_dir)
+
+  suf = get_suffix_number(out_dir)
+
+  model = models[process_name]
+
+  if process_name != 'cyanotype_full':
+      img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+
+  pred_img = model.predict_img(img)
+  out_path = f'{out_dir}/linear_{suf}.png'
+  cv2.imwrite(out_path, pred_img)
+
+  return model.predict_img(pred_img)
+
+
+def optimize_img(process_name, img):
+  out_dir = f'outputs/{process_name}'
+  if not os.path.exists(out_dir):
+    os.makedirs(out_dir)
+
+  suf = get_suffix_number(out_dir)
+
+  model = models[process_name]
+
+  if process_name != 'cyanotype_full':
+      img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+
+  x0, opt_img = model.tf_optimize(img)
+
+  out_path = f'{out_dir}/opt_{suf}.png'
+  cv2.imwrite(out_path, opt_img)
+  out_path = f'{out_dir}/x0_{suf}.png'
+  cv2.imwrite(out_path, x0)
+
+  return x0, opt_img

src/utils.py

@@ -0,0 +1,26 @@
+import numpy as np
+import cv2
+from PIL import Image
+
+
+def pil_to_cv(img):
+    new_img = np.array(img, dtype=np.uint8)
+    if new_img.ndim == 2:  # モノクロ
+        pass
+    elif new_img.shape[2] == 3:  # カラー
+        new_img = cv2.cvtColor(new_img, cv2.COLOR_RGB2BGR)
+    elif new_img.shape[2] == 4:  # 透過
+        new_img = cv2.cvtColor(new_img, cv2.COLOR_RGBA2BGRA)
+    return new_img
+
+def cv_to_pil(img):
+    new_img = img.copy()
+    if new_img.ndim == 2:  # モノクロ
+        pass
+    elif new_img.shape[2] == 3:  # カラー
+        new_img = cv2.cvtColor(new_img, cv2.COLOR_BGR2RGB)
+    elif new_img.shape[2] == 4:  # 透過
+        new_img = cv2.cvtColor(new_img, cv2.COLOR_BGRA2RGBA)
+    new_img = Image.fromarray(new_img)
+
+    return new_img