nihaomur / CNN_Practice

Exploring CNN-Based Model Construction through Practical Implementation

CNN Practice:

In this repository, I practiced building CNN models and later applied pre-trained models from the Keras API for Transfer Learning.

Minst:

Constructed a convolutional neural network for predicting the Keras-built MNIST dataset. Explored the use of image augmentation techniques, but observed a decrease in training accuracy for digits. Image augmentations, such as flipping, led to a loss of original discernible features.

Cat Dog Classify:

Utilized the pre-trained ResNet50 model to classify cat and dog images, achieving an accuracy of 97.06%.

Mango Level Classify:

Performed data augmentation on the dataset and trained using three pre-trained models: Xception, EfficientNet, and EfficientV2B2. Successfully addressed the mango level classification problem with accuracy of 85.25%.