Surface roughness prediction CNN

CNN for predicting surface roughness after milling machining process

Starting out

ctrl + shift + P -> create env, choose interpreter (latest python version)

Install libs

py -m pip install numpy

First moderately successful model

Overfitting is possible, for although val_accuracy=1.00, the model sometimes fails to predict correctly on new images

[INFO] ... [127/127] images were preprocessed successfully [INFO] Number of generated images: 1016 [INFO] Number of samples used for training ... 812 [INFO] Number of samples used for validation ... 204 25/25 [==============================] - 1364s 51s/step - loss: 1162.7634 - accuracy: 0.3573 - val_loss: 1.0552 - val_accuracy: 0.4706 [INFO] Current validation accuracy: 0.47058823704719543 Epoch 1/2 25/25 [==============================] - 656s 26s/step - loss: 1.1152 - accuracy: 0.4459 - val_loss: 0.9554 - val_accuracy: 0.5392 Epoch 2/2 25/25 [==============================] - 1276s 51s/step - loss: 0.9498 - accuracy: 0.5787 - val_loss: 0.7493 - val_accuracy: 0.8284 [INFO] Current validation accuracy: 0.8284313678741455 Epoch 1/3 25/25 [==============================] - 2660s 106s/step - loss: 0.5641 - accuracy: 0.8159 - val_loss: 0.3101 - val_accuracy: 0.9314 Epoch 2/3 25/25 [==============================] - 2721s 108s/step - loss: 0.1916 - accuracy: 0.9360 - val_loss: 0.0327 - val_accuracy: 1.0000 Epoch 3/3 25/25 [==============================] - 2640s 105s/step - loss: 0.0430 - accuracy: 0.9941 - val_loss: 0.0168 - val_accuracy: 1.0000 [INFO] Current validation accuracy: 1.0 [INFO] ... Saving surface roughness prediction model ... [INFO] The surface roughness prediction model was built SUCCESSFULLY [INFO] Run predictRoughness to test the model

Second moderately successful model (roughness_classifier5)

If both accuracy and validation accuracy are above 80% and the loss and validation loss are below 1, it generally indicates that your model is performing well on the task of predicting surface roughness based on images.

An accuracy above 80% suggests that model is making correct predictions for a significant portion of the dataset, which is a positive outcome. Similarly, having both loss and validation loss below 1 indicates that model's predictions are relatively close to the ground truth values, as the loss represents the discrepancy between predicted and actual values.

However, it's important to consider the specific context and requirements of your application. The definition of a "good" accuracy or loss depends on the problem at hand and the level of precision needed in predictions. In some scenarios, an accuracy of 80% might be acceptable, while in others, higher levels of accuracy may be required.

Since predicting surface roughness based on images, it's crucial to evaluate the model's performance on unseen data and consider other metrics or factors that might be relevant for specific use case. Additionally, it may be desirable to validate the model's performance on a separate test dataset to ensure that it generalizes well and is not overfitting to the training or validation data.

In summary, achieving an accuracy above 80%, having loss and validation loss below 1, and considering the context of surface roughness prediction from images are positive indications, but further evaluation and validation are recommended to ensure the model's effectiveness in practical scenarios.

[INFO] Number of loaded images: 254
Images saved successfully. [INFO] Number of generated images: 6096 [INFO] Splitting data into train and validation sets... [INFO] Number of samples used for training ... 4876 [INFO] Number of samples used for validation ... 1220 [INFO] Initialising the model training... Epoch 1/100 153/153 [==============================] - 182s 1s/step - loss: 8.7538 - accuracy: 0.4231 - val_loss: 3.6414 - val_accuracy: 0.4033 Epoch 2/100 153/153 [==============================] - 181s 1s/step - loss: 2.5569 - accuracy: 0.4969 - val_loss: 5.3112 - val_accuracy: 0.1811 Epoch 3/100 153/153 [==============================] - 178s 1s/step - loss: 1.8800 - accuracy: 0.5548 - val_loss: 2.5127 - val_accuracy: 0.4951 Epoch 4/100 153/153 [==============================] - 176s 1s/step - loss: 1.3119 - accuracy: 0.6458 - val_loss: 1.1386 - val_accuracy: 0.6516 Epoch 5/100 153/153 [==============================] - 176s 1s/step - loss: 0.9658 - accuracy: 0.6963 - val_loss: 1.7495 - val_accuracy: 0.6402 Epoch 6/100 153/153 [==============================] - 175s 1s/step - loss: 0.7730 - accuracy: 0.7533 - val_loss: 1.3607 - val_accuracy: 0.7164 Epoch 7/100 153/153 [==============================] - 175s 1s/step - loss: 0.6817 - accuracy: 0.7769 - val_loss: 1.0287 - val_accuracy: 0.7402 Epoch 8/100 153/153 [==============================] - 177s 1s/step - loss: 0.4855 - accuracy: 0.8310 - val_loss: 1.0270 - val_accuracy: 0.7869 Epoch 9/100 153/153 [==============================] - 175s 1s/step - loss: 0.4466 - accuracy: 0.8445 - val_loss: 0.8610 - val_accuracy: 0.8041 Epoch 10/100 153/153 [==============================] - 177s 1s/step - loss: 0.3455 - accuracy: 0.8753 - val_loss: 0.9185 - val_accuracy: 0.8254 Epoch 11/100 153/153 [==============================] - 175s 1s/step - loss: 0.3491 - accuracy: 0.8786 - val_loss: 1.0509 - val_accuracy: 0.8041 Epoch 12/100 153/153 [==============================] - 185s 1s/step - loss: 0.2968 - accuracy: 0.8999 - val_loss: 0.7661 - val_accuracy: 0.8369 Epoch 13/100 153/153 [==============================] - 178s 1s/step - loss: 0.2227 - accuracy: 0.9188 - val_loss: 0.8417 - val_accuracy: 0.8525 Epoch 14/100 153/153 [==============================] - 176s 1s/step - loss: 0.1868 - accuracy: 0.9374 - val_loss: 0.9051 - val_accuracy: 0.8533 Epoch 15/100 153/153 [==============================] - ETA: 0s - loss: 0.1814 - accuracy: 0.9354Restoring model weights from the end of the best epoch: 12. 153/153 [==============================] - 176s 1s/step - loss: 0.1814 - accuracy: 0.9354 - val_loss: 0.8084 - val_accuracy: 0.8648 Epoch 15: early stopping Model: "sequential"

---

Layer (type) Output Shape Param

conv2d (Conv2D) (None, 339, 339, 16) 160

batch_normalization (BatchN (None, 339, 339, 16) 64 ormalization)

max_pooling2d (MaxPooling2D (None, 169, 169, 16) 0 )

dropout (Dropout) (None, 169, 169, 16) 0

conv2d_1 (Conv2D) (None, 167, 167, 32) 4640

batch_normalization_1 (Batc (None, 167, 167, 32) 128 hNormalization)

max_pooling2d_1 (MaxPooling (None, 83, 83, 32) 0 2D)

dropout_1 (Dropout) (None, 83, 83, 32) 0

conv2d_2 (Conv2D) (None, 81, 81, 64) 18496

batch_normalization_2 (Batc (None, 81, 81, 64) 256 hNormalization)

max_pooling2d_2 (MaxPooling (None, 40, 40, 64) 0 2D)

dropout_2 (Dropout) (None, 40, 40, 64) 0

flatten (Flatten) (None, 102400) 0

dense (Dense) (None, 3) 307203

================================================================= Total params: 330,947 Trainable params: 330,723 Non-trainable params: 224

---

None [INFO] ... Saving surface roughness prediction model ... [INFO] The surface roughness prediction model was built SUCCESSFULLY [INFO] Plotting training statistics... [INFO] Plot completed... [INFO] Run predictRoughness.py to test the model