The following model represents a classification of the type of noise applied to a color or gray image.

Type of noises used are:

- Additive Gaussian noise
- Salt and pepper noise

Gray images: accuracy = 98%

- Architecture of models of convolution networks

We opted to use the LeNet-5 architecture. "LeNet-5 CNN architecture is made up of 7 layers. The layer composition consists of 3 convolutional layers, 2 subsampling layers and 2 fully connected layers."

- Result

We carried out the training of our model on a dataset of: - 38,000 images for training - 19000 images for validation Salt and pepper noise was applied to the training images by scaling (the percentage of noise is increased by 5% for every 1000 group of images to increase the training areas).

The results of the compilations with an epoch = 25 are as follows:

---

Epoch 20/25
113s 380ms/step - loss: 0.0064 - accuracy: 0.9978 - val_loss: 0.0409 - val_accuracy: 0.9901
Epoch 21/25
114s 383ms/step - loss: 0.0078 - accuracy: 0.9975 - val_loss: 0.0701 - val_accuracy: 0.9875
Epoch 22/25
113s 381ms/step - loss: 0.0040 - accuracy: 0.9987 - val_loss: 0.0965 - val_accuracy: 0.9837
Epoch 23/25
112s 378ms/step - loss: 0.0063 - accuracy: 0.9978 - val_loss: 0.0545 - val_accuracy: 0.9894
Epoch 24/25
113s 380ms/step - loss: 0.0075 - accuracy: 0.9974 - val_loss: 0.0607 - val_accuracy: 0.9883
Epoch 25/25
113s 379ms/step - loss: 0.0042 - accuracy: 0.9986 - val_loss: 0.0537 - val_accuracy: 0.9891

253s 3s/step - loss: 0.1215 - accuracy: 0.9811

---

- Testing and Experimentation

For the test phase we used a dataset of 9500 images and the classification result obtained is 100%. here are some examples of gray images with and without noise:

Grayscale image without noise

Grayscale image with salt & pepper noise

Color images: accuracy = 74%

- Architecture of models of convolution networks

We opted to use the LeNet-5 architecture. "LeNet-5 CNN architecture is made up of 7 layers. The layer composition consists of 3 convolutional layers, 2 subsampling layers and 2 fully connected layers."

- Result

We carried out the training of our model on a dataset of: - 57,000 images for training - 28,500 images for validation Salt and pepper and Gaussian noise were applied to the training images by scaling (the percentage of noise is increased by 5% for every 1000 group of images to increase the training areas).

The results of the compilations with epoch = 50 are as follows:

---

Epoch 45/50
126s 282ms/step - loss: 0.0331 - accuracy: 0.9879 - val_loss: 0.6799 - val_accuracy: 0.9207
Epoch 46/50
125s 281ms/step - loss: 0.0195 - accuracy: 0.9933 - val_loss: 1.8746 - val_accuracy: 0.9087
Epoch 47/50
126s 282ms/step - loss: 0.0266 - accuracy: 0.9905 - val_loss: 1.2452 - val_accuracy: 0.9162
Epoch 48/50
127s 284ms/step - loss: 0.0324 - accuracy: 0.9895 - val_loss: 1.5137 - val_accuracy: 0.9174
Epoch 49/50
126s 282ms/step - loss: 0.0149 - accuracy: 0.9948 - val_loss: 1.4898 - val_accuracy: 0.9112
Epoch 50/50
127s 284ms/step - loss: 0.0251 - accuracy: 0.9911 - val_loss: 1.9037 - val_accuracy: 0.9162

297s 3s/step - loss: 4.6600 - accuracy: 0.7484

---

- Testing and Experimentation

For the test phase we used a dataset of 14,250 images and the following results: - For the classification of images without noise: 99.8% degree of success - For the class of images with Gaussian noise: 86.7% - For the class of images with salt and pepper noise: 99.6%

here are some examples of gray images with and without noise:

Color image without noise

Color image with gaussian noise

Color image with salt & pepper noise