【Raspberry Pi】電子工作入門⑧～ディープラーニング入門～【Python】

入力画像（5×5）：
1 1 1 0 0
1 1 1 0 0
1 1 1 0 0
0 0 0 1 1
0 0 0 1 1

縦線フィルタ（3×3）：
1 0 -1
1 0 -1
1 0 -1

畳み込み結果（特徴マップ）：
3  3  -3 ...
3  3  -3 ...
...

入力（2×2の領域）：
1 3
2 8

MaxPooling → 8（最大値）

入力 (32×32 gray)
   ↓
Conv2D（3×3×8フィルタ）
   ↓
ReLU
   ↓
MaxPooling（2×2）
   ↓
Conv2D（3×3×16）
   ↓
ReLU
   ↓
MaxPooling（2×2）
   ↓
Flatten（1次元化）
   ↓
Dense（全結合）
   ↓
出力（3クラス：circle/square/triangle）

import tensorflow as tf
import numpy as np
import os
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras import layers, models

# ====== 1. 画像が入っているフォルダ ======
DATASET_DIR = "dataset"  # 前回作成済み

# ====== 2. データ読み込み（ImageDataGenerator） ======
datagen = ImageDataGenerator(
    rescale=1.0/255,
    validation_split=0.2       # 訓練80% / テスト20%
)

# 訓練データ
train_data = datagen.flow_from_directory(
    DATASET_DIR,
    target_size=(64, 64),       # CNN入力サイズ
    batch_size=8,
    class_mode='categorical',
    subset='training'
)

# 検証データ
val_data = datagen.flow_from_directory(
    DATASET_DIR,
    target_size=(64, 64),
    batch_size=8,
    class_mode='categorical',
    subset='validation'
)

# クラス名確認
print("クラス名（ラベル）:", train_data.class_indices)

# ====== 3. CNNモデル定義（小さめシンプル構造） ======
model = models.Sequential([
    layers.Conv2D(16, (3,3), activation='relu', input_shape=(64,64,3)),
    layers.MaxPooling2D(2,2),

    layers.Conv2D(32, (3,3), activation='relu'),
    layers.MaxPooling2D(2,2),

    layers.Flatten(),
    layers.Dense(64, activation='relu'),
    layers.Dense(3, activation='softmax')   # circle / square / triangle
])

model.compile(
    optimizer='adam',
    loss='categorical_crossentropy',
    metrics=['accuracy']
)

model.summary()

# ====== 4. 学習 ======
history = model.fit(
    train_data,
    epochs=10,
    validation_data=val_data
)

# ====== 5. 学習済みモデル保存 ======
model.save("shape_cnn.h5")
print("shape_cnn.h5 を保存しました！")

# ====== 6. TFLite 変換（PC側のみ成功します） ======
converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()

with open("shape_cnn.tflite", "wb") as f:
    f.write(tflite_model)

print("shape_cnn.tflite を作成しました！")

import tensorflow as tf
import numpy as np
import os
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras import layers, models

DATASET_DIR = "dataset"

datagen = ImageDataGenerator(
    rescale=1.0/255,
    validation_split=0.2
)

train_data = datagen.flow_from_directory(
    DATASET_DIR,
    target_size=(64, 64),
    batch_size=8,
    class_mode='categorical',
    subset='training'
)

val_data = datagen.flow_from_directory(
    DATASET_DIR,
    target_size=(64, 64),
    batch_size=8,
    class_mode='categorical',
    subset='validation'
)

print("クラス名（ラベル）:", train_data.class_indices)

model = models.Sequential([
    layers.Conv2D(16, (3,3), activation='relu', input_shape=(64,64,3)),
    layers.MaxPooling2D(2,2),

    layers.Conv2D(32, (3,3), activation='relu'),
    layers.MaxPooling2D(2,2),

    layers.Flatten(),
    layers.Dense(64, activation='relu'),
    layers.Dense(3, activation='softmax')
])

model.compile(
    optimizer='adam',
    loss='categorical_crossentropy',
    metrics=['accuracy']
)

history = model.fit(
    train_data,
    epochs=10,
    validation_data=val_data
)

model.save("shape_cnn.h5")
print("shape_cnn.h5 を保存しました！")

converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()

with open("shape_cnn.tflite", "wb") as f:
    f.write(tflite_model)

pip3 install --break-system-packages tflite-runtime

from picamera2 import Picamera2
import cv2
import numpy as np
import tflite_runtime.interpreter as tflite
import time

# ===== 1. TFLite モデルの読み込み =====
MODEL_PATH = "shape_cnn.tflite"

interpreter = tflite.Interpreter(model_path=MODEL_PATH)
interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

# 学習時のクラス名（PC側の train_cnn.py と対応）
CLASS_NAMES = ["circle", "square", "triangle"]


def main():
    # カメラ初期化
    cam = Picamera2()
    cam.configure(cam.create_preview_configuration(main={"size": (640, 480)}))
    cam.start()

    # デジタルズームを解除
    try:
        sw, sh = cam.camera_properties["PixelArraySize"]
        cam.set_controls({"ScalerCrop": (0, 0, int(sw), int(sh))})
    except Exception:
        # うまく取得できない場合は、OV5647 の典型的なフル解像度を指定
        cam.set_controls({"ScalerCrop": (0, 0, 2592, 1944)})

    print("q：終了")

    while True:
        # カメラ画像を取得（RGB→BGR）
        frame = cv2.cvtColor(cam.capture_array(), cv2.COLOR_RGB2BGR)
        h, w, _ = frame.shape

        #  画面中央にガイド枠（ここだけを分類に使う）
        gx1, gy1 = w // 4, h // 4
        gx2, gy2 = w * 3 // 4, h * 3 // 4
        cv2.rectangle(frame, (gx1, gy1), (gx2, gy2), (0, 255, 255), 2)
        cv2.putText(frame, "Put object inside box",
                    (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 255, 255), 2)

        # ガイド枠内の画像を切り出し
        roi = frame[gy1:gy2, gx1:gx2]

        #CNN入力用に前処理
        # 64x64 にリサイズ（学習時と同じサイズ）
        img = cv2.resize(roi, (64, 64))
        # 0〜1に正規化
        img = img.astype(np.float32) / 255.0
        # (1, 64, 64, 3) に形を揃える
        img = np.expand_dims(img, axis=0)

        #TFLite 推論
        interpreter.set_tensor(input_details[0]["index"], img)
        interpreter.invoke()
        output = interpreter.get_tensor(output_details[0]["index"])[0]

        pred_id = int(np.argmax(output))
        label = CLASS_NAMES[pred_id]
        prob = float(output[pred_id])

        # 結果を画面左上に表示
        cv2.putText(frame, f"{label} ({prob:.2f})",
                    (10, 70), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 255, 0), 2)

        # そのまま表示
        cv2.imshow("CNN Camera", frame)

        key = cv2.waitKey(1) & 0xFF
        if key == ord("q"):
            break

    # ===== 8. 終了処理 =====
    cv2.destroyAllWindows()
    cam.stop()


if __name__ == "__main__":
    main()

from picamera2 import Picamera2
import cv2
import numpy as np
import tflite_runtime.interpreter as tflite
import time

MODEL_PATH = "shape_cnn.tflite"

interpreter = tflite.Interpreter(model_path=MODEL_PATH)
interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

CLASS_NAMES = ["circle", "square", "triangle"]

cam = Picamera2()
cam.configure(cam.create_preview_configuration(main={"size": (640, 480)}))
cam.start()

try:
    sw, sh = cam.camera_properties["PixelArraySize"]
    cam.set_controls({"ScalerCrop": (0, 0, int(sw), int(sh))})
except Exception:
    cam.set_controls({"ScalerCrop": (0, 0, 2592, 1944)})

frame = cv2.cvtColor(cam.capture_array(), cv2.COLOR_RGB2BGR)
h, w, _ = frame.shape

gx1, gy1 = w // 4, h // 4
gx2, gy2 = w * 3 // 4, h * 3 // 4
cv2.rectangle(frame, (gx1, gy1), (gx2, gy2), (0, 255, 255), 2)
cv2.putText(frame, "Put object inside box",
            (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 255, 255), 2)

roi = frame[gy1:gy2, gx1:gx2]

img = cv2.resize(roi, (64, 64))
img = img.astype(np.float32) / 255.0
img = np.expand_dims(img, axis=0)

interpreter.set_tensor(input_details[0]["index"], img)
interpreter.invoke()
output = interpreter.get_tensor(output_details[0]["index"])[0]

pred_id = int(np.argmax(output))
label = CLASS_NAMES[pred_id]
prob = float(output[pred_id])

cv2.putText(frame, f"{label} ({prob:.2f})",
            (10, 70), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 255, 0), 2)

cv2.imshow("CNN Camera", frame)

key = cv2.waitKey(1) & 0xFF
if key == ord("q"):
    break

cv2.destroyAllWindows()
cam.stop()