Modern Convolutional Neural Networks

What is modern convolutional neural networks?

Modern convolutional neural networks (CNNs) have revolutionized the field of computer vision by achieving remarkable performance in various tasks. These deep learning models are specifically designed to process visual data, mimicking the human visual system. Through their hierarchical architecture, CNNs effectively learn and extract increasingly complex features from input images, enabling them to identify patterns and objects with exceptional accuracy. Modern CNNs incorporate advanced techniques such as residual connections, attention mechanisms, and normalization layers, enhancing their ability to handle large-scale datasets and address challenges like overfitting. Their widespread adoption has led to breakthroughs in image classification, object detection, semantic segmentation, and other visual recognition tasks, making CNNs an indispensable tool in modern artificial intelligence research and applications.

About VGG network?

The VGG algorithm, short for the Visual Geometry Group algorithm, is a deep convolutional neural network architecture that has proven to be highly influential in the field of computer vision. Proposed by the Visual Geometry Group at the University of Oxford, the VGG algorithm consists of a series of convolutional layers followed by fully connected layers. What sets VGG apart is its uniform architecture, where convolutional layers have a small receptive field of 3×3 and are stacked on top of each other, making the network deeper. This depth allows VGG to learn increasingly complex features, leading to improved accuracy in image classification tasks. Despite its simplicity compared to more recent architectures, VGG has been widely adopted and serves as a strong baseline for benchmarking new models in the field of deep learning.

About this project

Many of my research projects focus on deep learning algorithms specifically designed for modern convolutional neural networks . For instance, at beginning, I have worked on 3,278 instances (5 classes) of apples for image processing and classification .

I use these libraries for classification of a dataset of apples:

Keras, Sklearn, pandas, TensorFlow, glob, Numpy and ImageDataGenerator.

Additionally, I have utilized these conditions for image data #augmentation, such as rotation or shifting:

“20 degrees for rotation and 10 percent for shift at x or y.”

Data splitted to 2450 for the train and 837 for the test phase at 5 classes.

I have employed the VGG and Inception  (InceptionV3) networks for data classification within the context of modern convolutional neural networks.

Results

The results obtained from these experiments are presented below:

• VGG accuracy: 0.9943
• VGG val_accuracy: 0.8996
• Inception accuracy: 0.9861  
• Inception val_accuracy: 0.9367