<aside> ☝🏻

This Lab is based on this book: https://mlsysbook.ai/book/ More specifically, on this chapter → https://mlsysbook.ai/kits/contents/seeed/xiao_esp32s3/image_classification/image_classification.html

</aside>

Overview

We are increasingly facing an artificial intelligence (AI) revolution, where Edge AI and Computer Vision have a very high impact potential, and it is for now!

When we examine Machine Learning (ML) applied to vision, the first concept that greets us is Image Classification, a kind of ML’s Hello World that is both simple and profound.

The XIAO ESP32-S3 Sense, featuring an integrated OV3660 camera and SD card support, makes the XIAO ESP32-S3 Sense an excellent starting point for exploring TinyML vision AI.

In this Lab, we will explore Image Classification using the non-code tool SenseCraft AI and delve into more detailed development with Edge Impulse Studio and the Arduino IDE.

Learning Objectives

• Deploy Pre-trained Models using SenseCraft AI Studio for immediate computer vision applications
• Collect and Manage Image Datasets for custom classification tasks with proper data organization
• Train Custom Image Classification Models using transfer learning with MobileNet V2 architecture
• Optimize Models for Edge Deployment through quantization and memory-efficient preprocessing
• Implement Post-processing Pipelines, including GPIO control and real-time inference integration
• Compare Development Approaches between no-code and advanced ML platforms for embedded applications

Image Classification

Image classification is a fundamental task in computer vision that categorizes images into predefined classes. This process involves analyzing an image's visual content and assigning it a label from a fixed set of categories based on the dominant object or scene it depicts.

Image classification is crucial across a range of applications, from organizing and searching large image databases in digital libraries and social media platforms to enabling autonomous systems to comprehend their surroundings.

For example, image classification is widely used in rural contexts to support agriculture, environment, and community development. Key application areas include crop disease detection, pest identification, weed classification for precision farming, livestock health monitoring, soil and land-cover classification, forest and deforestation monitoring, quality grading of agricultural produce, water resource monitoring, and rural infrastructure mapping. These uses enable farmers to make informed decisions, enhance productivity, lower costs, and promote sustainable rural development.

Common architectures that have significantly advanced image classification include Convolutional Neural Networks (CNNs), such as AlexNet, VGGNet, and ResNet. These models have demonstrated remarkable accuracy on challenging datasets, such as ImageNet, by learning hierarchical representations of visual data.

As the cornerstone of many computer vision systems, image classification drives innovation, laying the groundwork for more complex tasks such as object detection and image segmentation, and facilitating a deeper understanding of visual data across industries.

<aside> 📌

We will explore the Person Classification model (“Person - No Person”), a ready-to-use computer vision application on the SenseCraft AI.

</aside>