All case studies

Computer visionResearch portfolio project

Classify a bird image at the order level

This experiment tests image classification across a dataset labelled with 42 bird orders, comparing model runs and reviewing images before inference.

Contribution
Mehrab Mashrafi
Scope
Image cleaning, transfer learning and inference
Project context
Original portfolio experiment

Classification scope

42 order labels

An order is a taxonomic group. This is not identification of every bird species.

Choose the level of identification

Grouping images by bird order makes this a different task from naming an individual species. Images were collected using different poses and search terms, then checked for unsuitable or misplaced examples.

Inspecting high loss images helped identify errors in the training material.

Keep comparisons tied to their evaluation

The original project compares ResNet, DenseNet and VGG runs and describes a selected ResNet model. Results from different runs should not become a single headline accuracy.

The source notebooks allow a reader to inspect preparation and evaluation. A separate test set would be needed before claiming performance on new field images.

How the work connects

  1. 01

    Order labels

  2. 02

    Image collection

  3. 03

    Loss review & cleaning

  4. 04

    Model comparison

  5. 05

    Order prediction

What this work establishes

A documented image collection and model comparison support an order classification demonstration, with notebooks available to inspect.

This is an order classification experiment. Web imagery, label quality and dataset splits affect its results; independent field performance is not established.

Inspect the work or discuss the approach

Have a similar workflow to improve?

Tell us who uses it, which records it depends on and where the handoff slows down. We can discuss a practical implementation scope.

Discuss your use case

Project notes from Mehrab Mashrafi and AI Byters. Updated .