Image-Classification-using-CNN

CIFAR-10 Dataset

The CIFAR-10 (Canadian Institute for Advanced Research) dataset is a widely used benchmark dataset in computer vision and deep learning. It consists of 60,000 color images categorized into 10 classes, with 6,000 images per class. The dataset is divided into:

• 45,000 training images
• 5,000 validation images
• 10,000 test images

Each image is 32×32 pixels with 3 color channels (RGB). The dataset is often used for training and evaluating image classification models.

Classes in CIFAR-10

The dataset includes the following 10 object categories:

Image Source: Papers With Code

CNN Model

The CNN model built for classification task of CIFAR-10 dataset has the following architecture:

Layer Type	Filters	Kernel Size	Activation	Other Features
Conv2D	32	(3,3)	ReLU	L2 Regularization, BatchNorm
Conv2D	32	(3,3)	ReLU	L2 Regularization, BatchNorm
MaxPooling2D	-	(2,2)	-	Reduces spatial size
Dropout	-	-	-	40% dropout to prevent overfitting
Conv2D	64	(3,3)	ReLU	L2 Regularization, BatchNorm
Conv2D	64	(3,3)	ReLU	L2 Regularization, BatchNorm
MaxPooling2D	-	(2,2)	-	Downsampling
Dropout	-	-	-	40% dropout
Conv2D	128	(3,3)	ReLU	L2 Regularization, BatchNorm
Conv2D	128	(3,3)	ReLU	L2 Regularization, BatchNorm
MaxPooling2D	-	(2,2)	-	Downsampling
Dropout	-	-	-	50% dropout
Conv2D	256	(3,3)	ReLU	L2 Regularization, BatchNorm
Conv2D	256	(3,3)	ReLU	L2 Regularization, BatchNorm
MaxPooling2D	-	(2,2)	-	Downsampling
Dropout	-	-	-	50% dropout
Flatten	-	-	-	Converts 2D features to 1D
Dense	256	-	ReLU	L2 Regularization, Dropout (30%)
Dense	10	-	Softmax	Final classification layer

Model Performance & Generalization

• With a test accuracy of 88.49%, our model demonstrates strong generalization to unseen data. This balance between accuracy and efficiency highlights its ability to learn meaningful patterns rather than just memorizing the training data.
• The model's effectiveness was further validated by testing on 5 unseen images sourced from Google. Impressively, it correctly classified all 5 images, demonstrating its strong generalization beyond the training dataset.