Introduction

This topic systematically addresses the foundations and applications of computer vision systems based on deep learning. The objective is to understand how to process and model visual information, both in color image format (RGB format) and grayscale, and how these representation decisions influence the models' ability to learn relevant patterns.

First, different image representation formats and their impact on subsequent processing are studied. The discussion analyzes when it is convenient to work with color images, leveraging chromatic information, and when it is preferable to convert them to grayscale to simplify the problem, reduce dimensionality, or focus attention on luminance structures. The advantages and limitations of each approach and their relationship to the type of task (classification, segmentation, detection, etc.) are discussed.

Next, various image processing techniques closely related to deep learning and, in many cases, signal processing are introduced. The section explains how to normalize input data, both at the image and dataset level, and analyzes the influence of these normalizations on numerical stability, convergence speed, and final model performance. On this basis, normalization mechanisms applied within the architectures themselves are also studied, such as Batch Normalization and Layer Normalization, comparing their operating principles, their implicit assumptions, and their effect when normalized based on data statistics versus the use of internal model parameters.

A central block of the topic is dedicated to the study of convolutional layers. The discussion shows how to implement a convolution from scratch, analyzing in detail the filter sliding operation over the image, the local linear combination of pixels, and the obtaining of feature maps. Subsequently, the section explains how to leverage primitives from libraries like PyTorch to define and train convolutional layers efficiently, maintaining intuition about what is happening at the mathematical and computational level.

In relation to convolutions, properties such as translational invariance and their practical limitations are discussed with special attention. Although the ideal convolution presents certain translational invariance, this property is altered when strides greater than 1 or pooling layers are introduced. In particular, the discussion analyzes how shifts (even circular ones) of the same image can produce significantly different responses due to these mechanisms, and reflects on why, despite this, pooling and stride are necessary to control spatial resolution, reduce computational cost, and increase the effective receptive field of deep layers.

It is also emphasized that, even when convolutional layers do not require a fixed input size, image resolution significantly conditions the type of features the network is capable of learning. The discussion addresses how the distribution of spatial frequencies and the scale of present patterns (edges, textures, global structures) influence the model's sensitivity to different levels of detail. This is essential to understand why the same architecture can behave differently depending on the size and quality of input images.

From a historical perspective, the topic covers some of the most influential convolutional architectures. It starts with LeNet, developed by Yann LeCun and his team, used in a pioneering way in production environments for handwritten digit recognition in bank checks, and continues with subsequent architectures that explore the systematic increase in network depth and width. Models like VGG, which deepens the use of homogeneous convolutional blocks, and ResNet, which introduces residual connections to facilitate training of very deep networks, are analyzed. This review allows contextualizing progress in computer vision and understanding the motivations behind the design of increasingly complex networks.

On this basis, transfer learning techniques are studied, which allow reusing convolutional models pretrained by large organizations (such as Meta, Google, or others) on massive datasets. The section shows how to adapt these models to specific problems through fine-tuning of the last layers, partial freezing of parameters, or redefinition of classification heads, thus reducing training cost and improving performance in scenarios with limited data.

In relation to efficiency and deployment on resource-constrained devices, methods such as knowledge distillation are introduced, which consist of transferring the behavior of large models (teachers) to smaller models (students). The discussion addresses how these techniques allow compressing convolutional networks and adapting them to embedded environments, such as mobile phones, microcontrollers, or other edge computing systems, while maintaining a significant fraction of their predictive capacity.

The topic also addresses autoencoders, both dense and convolutional, as tools for unsupervised learning of representations from images. The discussion analyzes how these architectures allow extracting useful latent features for compression, anomaly detection, and out-of-distribution sample detection. The basic encoder-decoder structure, the reconstruction function, and the interpretation of latent variables are explained.

At a more advanced level, some attention mechanisms applied to convolutional networks are presented, with special emphasis on channel attention. These mechanisms modulate the relative importance of each feature map, allowing the network to focus on spatially relevant patterns and enriching the extracted information without collapsing representations. The discussion addresses how these attention blocks can be integrated into standard convolutional architectures and what benefits they provide in terms of performance and robustness.

Subsequently, the application of Transformer-type architectures to the visual domain is introduced, through models like Vision Transformer (ViT). The central idea of dividing the image into patches, projecting them into an embedding space, and applying self-attention mechanisms similar to those used in natural language processing is explained. The original work proposed by Google is analyzed, as well as the implications of transferring the global attention paradigm to the computer vision context.

Finally, the field of explainability and interpretability of vision models is addressed, with the aim of understanding why a model makes certain classification decisions. Techniques based on weights and gradients are explored, as well as the generation of heat maps that indicate the most relevant image regions for prediction. Among these techniques is Grad-CAM and related methods, which allow visualizing the effective attention of deep layers on the input image. These tools are essential for auditing models, detecting biases, debugging unexpected behaviors, and increasing confidence in systems deployed in sensitive environments.