INR-Currency-Note-Recognition.ipynb

{"cells":[{"cell_type":"markdown","source":["# INDIAN RUPEES CURRENCY NOTE DENOMINATION RECOGNITION USING TRANSFER LEARNING WITH TENSORFLOW\n","\n","\n","---\n","\n","\n","* This project is a Convolutional Neural Network (CNN) based image classifier model\n","* Due to CNN models not performing well when built from scratch, this project makes use of feature extraction and fine tuning from pretrained CNNs\n","* Project developed using Tensorflow Keras\n","* Dataset sourced from [Kaggle](https://www.kaggle.com/code/kerneler/starter-indian-currency-note-images-f2ab9d26-d/data) "],"metadata":{"id":"TXuydp5_5WFX"}},{"cell_type":"code","execution_count":null,"metadata":{"id":"ZSSOc7lc1PN_"},"outputs":[],"source":["import tensorflow as tf"]},{"cell_type":"code","source":["from google.colab import drive\n","drive.mount('/content/drive')"],"metadata":{"colab":{"base_uri":"https://localhost:8080/"},"id":"319fkZ-EH5Ck","executionInfo":{"status":"ok","timestamp":1653621155687,"user_tz":-330,"elapsed":21399,"user":{"displayName":"Kailash N","userId":"17040189032999445576"}},"outputId":"b8bc65c9-65fb-47ff-c520-02966b81ae74"},"execution_count":null,"outputs":[{"output_type":"stream","name":"stdout","text":["Mounted at /content/drive\n"]}]},{"cell_type":"markdown","source":["\n","\n","---\n","\n","\n","# Dataset Preparation\n","\n","\n","---\n","\n","\n","* The dataset does not contain the test images segregated by folders for different classes. So it is not possible to derive the class labels for the test images using Keras image_dataset_from_directory utility\n","* For this purpose, the images were manually moved to separate folders corresponding to each class of the image"],"metadata":{"id":"UhWvif-_1POE"}},{"cell_type":"code","execution_count":null,"metadata":{"id":"5e4BFEJC1POF"},"outputs":[],"source":["train_dir = '/content/drive/MyDrive/image_classification/final/training'\n","test_dir ='/content/drive/MyDrive/image_classification/final/test'\n","validation_dir = '/content/drive/MyDrive/image_classification/final/validation'"]},{"cell_type":"code","execution_count":null,"metadata":{"id":"USLG_lg_1POG","outputId":"9f26bfe4-cea0-4fe4-8d3b-86008d6b2719","colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"status":"ok","timestamp":1653623497147,"user_tz":-330,"elapsed":517,"user":{"displayName":"Kailash N","userId":"17040189032999445576"}}},"outputs":[{"output_type":"stream","name":"stdout","text":["Found 3582 files belonging to 8 classes.\n","Found 95 files belonging to 8 classes.\n","Found 345 files belonging to 8 classes.\n"]}],"source":["from tensorflow.keras.utils import image_dataset_from_directory\n","\n","tf.random.set_seed(42)\n","\n","train_data= image_dataset_from_directory(train_dir, label_mode = 'categorical', shuffle=True, seed=42)\n","test_data = image_dataset_from_directory(test_dir, label_mode = 'categorical', shuffle=True)\n","validation_data = image_dataset_from_directory(validation_dir, label_mode = 'categorical', shuffle=True)"]},{"cell_type":"markdown","source":["\n","\n","---\n","\n","\n","# Data Augmentation Layer\n","\n","\n","---\n","\n","\n","* Can add and experiment around with rescaling, random zoom, random height, etc.\n","* Rescaling must be avoided for EfficientNetB0"],"metadata":{"id":"yn5coDIt8VMT"}},{"cell_type":"code","execution_count":null,"metadata":{"id":"iZtoHfpM1POF"},"outputs":[],"source":["# Data Augmentation/Preprocessing\n","from tensorflow import keras\n","from tensorflow.keras.layers.experimental import preprocessing\n","\n","data_augmentation_layer = keras.Sequential([\n","    preprocessing.Rescaling(1/255.) #should not be used with efficientnet\n","])"]},{"cell_type":"markdown","source":["\n","\n","---\n","\n","\n","# Transfer Learning\n","\n","\n","---\n","\n","\n","* Transfer learning using various pretrained models can be performed for results analysis comparisom\n","\n","* Model(s) completed till now :\n","  * EfficientNetB0"],"metadata":{"id":"0qL7cD106-Da"}},{"cell_type":"markdown","source":["## EfficientNetB0\n","\n","\n","---\n","\n"],"metadata":{"id":"KTtILn_c7IEC"}},{"cell_type":"markdown","source":["### Feature Extraction - Create"],"metadata":{"id":"Fd2vIyry7XqJ"}},{"cell_type":"code","execution_count":null,"metadata":{"id":"JwzOqXhe1POH"},"outputs":[],"source":["from tensorflow.keras import layers\n","from tensorflow.keras.applications import EfficientNetB0\n","\n","tf.random.set_seed(42)\n","base = EfficientNetB0(include_top = False)\n","base.trainable = False\n","\n","inputs = layers.Input(shape=(256,256,3))\n","\n","#x = data_augmentation_layer(inputs)\n","\n","x = base(inputs, training = False)\n","\n","x = layers.GlobalAveragePooling2D()(x)\n","\n","outputs = layers.Dense(8, activation=\"softmax\")(x)\n","\n","efficient_net_b0 = tf.keras.Model(inputs, outputs)"]},{"cell_type":"markdown","source":["### Feature Extraction - Compile"],"metadata":{"id":"yqqP-yUO7aaK"}},{"cell_type":"code","execution_count":null,"metadata":{"id":"u9i6Vlw11POI"},"outputs":[],"source":["# compile\n","efficient_net_b0.compile(loss=\"categorical_crossentropy\",optimizer=tf.keras.optimizers.Adam(),metrics=[\"accuracy\"])"]},{"cell_type":"markdown","source":["### Feature Extraction - Fit"],"metadata":{"id":"uVvOM_G07c1C"}},{"cell_type":"code","execution_count":null,"metadata":{"id":"3Rm_xLYd1POJ","outputId":"ec0e31d6-7dd9-48cd-c9da-d42921435efe","colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"status":"ok","timestamp":1653624081288,"user_tz":-330,"elapsed":552148,"user":{"displayName":"Kailash N","userId":"17040189032999445576"}}},"outputs":[{"output_type":"stream","name":"stdout","text":["Epoch 1/7\n","112/112 [==============================] - 85s 631ms/step - loss: 1.3790 - accuracy: 0.5723 - val_loss: 1.0000 - val_accuracy: 0.7344\n","Epoch 2/7\n","112/112 [==============================] - 70s 588ms/step - loss: 0.8188 - accuracy: 0.7929 - val_loss: 0.7270 - val_accuracy: 0.8125\n","Epoch 3/7\n","112/112 [==============================] - 70s 591ms/step - loss: 0.6205 - accuracy: 0.8587 - val_loss: 0.4898 - val_accuracy: 0.9219\n","Epoch 4/7\n","112/112 [==============================] - 69s 584ms/step - loss: 0.5098 - accuracy: 0.8900 - val_loss: 0.5681 - val_accuracy: 0.8594\n","Epoch 5/7\n","112/112 [==============================] - 70s 586ms/step - loss: 0.4357 - accuracy: 0.9068 - val_loss: 0.4386 - val_accuracy: 0.8906\n","Epoch 6/7\n","112/112 [==============================] - 70s 588ms/step - loss: 0.3801 - accuracy: 0.9196 - val_loss: 0.3120 - val_accuracy: 0.9844\n","Epoch 7/7\n","112/112 [==============================] - 69s 585ms/step - loss: 0.3390 - accuracy: 0.9299 - val_loss: 0.3782 - val_accuracy: 0.9062\n"]}],"source":["#fit\n","history_efficientnetb0 = efficient_net_b0.fit(train_data, epochs = 7 , validation_data=validation_data, validation_steps = int(0.25*len(validation_data)))"]},{"cell_type":"code","source":["efficient_net_b0.evaluate(test_data)"],"metadata":{"id":"1TY9LdvbtLhQ","colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"status":"ok","timestamp":1653624083894,"user_tz":-330,"elapsed":2614,"user":{"displayName":"Kailash N","userId":"17040189032999445576"}},"outputId":"0d3665f7-713f-40fd-9f13-b4f023cdaf8a"},"execution_count":null,"outputs":[{"output_type":"stream","name":"stdout","text":["3/3 [==============================] - 3s 363ms/step - loss: 0.3324 - accuracy: 0.9474\n"]},{"output_type":"execute_result","data":{"text/plain":["[0.33238035440444946, 0.9473684430122375]"]},"metadata":{},"execution_count":8}]},{"cell_type":"markdown","source":["### Fine Tuning\n"],"metadata":{"id":"uZaCBtGUsqxt"}},{"cell_type":"code","source":["#Last 10 layers are trainable\n","base.trainable=True\n","\n","for layers in base.layers[:-5]:\n","  layers.trainable=False"],"metadata":{"id":"gnSYo6WwszcN"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["checkpoint_dir = '/content/drive/MyDrive/image_classification/final/efficient_net_b0.h5'\n","\n","cp_callback = tf.keras.callbacks.ModelCheckpoint(checkpoint_dir,save_weights_only=True,save_best_only=True)\n","\n","efficient_net_b0.compile(loss=tf.keras.losses.CategoricalCrossentropy(),\n","                         optimizer=tf.keras.optimizers.Adam(0.0001), \n","                        metrics=[\"accuracy\"])"],"metadata":{"id":"5FzkdXoruaZT"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["efficient_net_b0.fit(train_data, epochs = 15, initial_epoch=7, validation_data = validation_data, validation_steps = int(0.15*len(validation_data)),\n","                     callbacks=[cp_callback])"],"metadata":{"colab":{"base_uri":"https://localhost:8080/"},"id":"oNpySqJlu7SM","executionInfo":{"status":"ok","timestamp":1653624675446,"user_tz":-330,"elapsed":591568,"user":{"displayName":"Kailash N","userId":"17040189032999445576"}},"outputId":"ee64bac4-72f5-4172-8217-995db694a22d"},"execution_count":null,"outputs":[{"output_type":"stream","name":"stdout","text":["Epoch 8/15\n","112/112 [==============================] - 77s 605ms/step - loss: 0.2642 - accuracy: 0.9361 - val_loss: 0.2005 - val_accuracy: 0.9375\n","Epoch 9/15\n","112/112 [==============================] - 70s 589ms/step - loss: 0.1695 - accuracy: 0.9604 - val_loss: 0.1433 - val_accuracy: 0.9375\n","Epoch 10/15\n","112/112 [==============================] - 69s 584ms/step - loss: 0.1293 - accuracy: 0.9743 - val_loss: 0.2275 - val_accuracy: 0.9062\n","Epoch 11/15\n","112/112 [==============================] - 71s 597ms/step - loss: 0.0972 - accuracy: 0.9805 - val_loss: 0.1139 - val_accuracy: 0.9688\n","Epoch 12/15\n","112/112 [==============================] - 70s 592ms/step - loss: 0.0783 - accuracy: 0.9886 - val_loss: 0.2841 - val_accuracy: 0.9375\n","Epoch 13/15\n","112/112 [==============================] - 71s 599ms/step - loss: 0.0645 - accuracy: 0.9897 - val_loss: 0.0563 - val_accuracy: 0.9688\n","Epoch 14/15\n","112/112 [==============================] - 70s 593ms/step - loss: 0.0532 - accuracy: 0.9913 - val_loss: 0.0779 - val_accuracy: 0.9688\n","Epoch 15/15\n","112/112 [==============================] - 71s 594ms/step - loss: 0.0472 - accuracy: 0.9927 - val_loss: 0.1808 - val_accuracy: 0.9062\n"]},{"output_type":"execute_result","data":{"text/plain":["<keras.callbacks.History at 0x7f122426eb90>"]},"metadata":{},"execution_count":11}]},{"cell_type":"code","source":["efficient_net_b0.evaluate(test_data)"],"metadata":{"colab":{"base_uri":"https://localhost:8080/"},"id":"sCDE8H8Mgg2G","executionInfo":{"status":"ok","timestamp":1653624676880,"user_tz":-330,"elapsed":1440,"user":{"displayName":"Kailash N","userId":"17040189032999445576"}},"outputId":"0a36faff-ffb2-48fa-fdae-22fec40dff22"},"execution_count":null,"outputs":[{"output_type":"stream","name":"stdout","text":["3/3 [==============================] - 2s 95ms/step - loss: 0.0313 - accuracy: 1.0000\n"]},{"output_type":"execute_result","data":{"text/plain":["[0.03130357712507248, 1.0]"]},"metadata":{},"execution_count":12}]},{"cell_type":"code","source":["efficient_net_b0.load_weights(checkpoint_dir)\n","efficient_net_b0.evaluate(test_data)"],"metadata":{"colab":{"base_uri":"https://localhost:8080/"},"id":"EZe0PIm4t8Pd","executionInfo":{"status":"ok","timestamp":1653624900666,"user_tz":-330,"elapsed":3861,"user":{"displayName":"Kailash N","userId":"17040189032999445576"}},"outputId":"35308a17-5163-4a1c-bcde-c90c59cd7af6"},"execution_count":null,"outputs":[{"output_type":"stream","name":"stdout","text":["3/3 [==============================] - 2s 99ms/step - loss: 0.0374 - accuracy: 1.0000\n"]},{"output_type":"execute_result","data":{"text/plain":["[0.037449274212121964, 1.0]"]},"metadata":{},"execution_count":13}]}],"metadata":{"interpreter":{"hash":"8948cf1f0e8fea22eeb78dc4968d83690bfca6b4573734f97156567b480a831c"},"kernelspec":{"display_name":"Python 3.8.10 ('prism')","language":"python","name":"python3"},"language_info":{"codemirror_mode":{"name":"ipython","version":3},"file_extension":".py","mimetype":"text/x-python","name":"python","nbconvert_exporter":"python","pygments_lexer":"ipython3","version":"3.8.10"},"orig_nbformat":4,"colab":{"name":"INR_recognition.ipynb","provenance":[],"collapsed_sections":["0qL7cD106-Da"]},"accelerator":"GPU"},"nbformat":4,"nbformat_minor":0}