Insect Pest Classification with ViT, DeiT & Ensemble Fusion

Introduction

Insect pests destroy a large share of the world's crops every year, pushing up pesticide use, raising costs for farmers, and putting food supply at risk. Spotting pests by eye is slow, needs expert knowledge, and simply cannot cover vast farmland.

Image software offers a faster path: it can recognise harmful insects from ordinary photos, helping catch a problem early before it spreads. This project builds a system that identifies crop-damaging pests from pictures.

Problem Statement

Relying on people to spot pests runs into real limits:

1. Many species look alike: especially across life stages and lighting, so they get mixed up.
2. One expert cannot possibly watch thousands of acres at once.
3. Some pests are common while rarer: damaging ones show up far less in the data.
4. Field tools need to run on a phone or small device: not only powerful computers.
5. Late detection often leads to heavy: blanket spraying instead of targeted treatment.

Objectives

1. Identify several high-impact crop pests from images.
2. Use modern image models and adapt them to pest photos.
3. Combine several models so their strengths add up.
4. Produce a lighter version that can run on everyday devices.
5. Handle the fact that some pests appear far more often than others.

Dataset

The work is built around a collection of labelled pest images. Depending on the client, this can be a public agricultural benchmark, licensed or private field photos, images from other trusted sources, or a custom set gathered for the project. Images are cleaned, resized, balanced across pest types, and lightly varied so the system copes with real-world conditions.

Methodology

The project adapts ViT and DeiT to recognise pests, then combines their outputs with early fusion, late fusion, and majority voting so the group decision is steadier than any single model. A lighter student version based on EfficientNet-B4 is also trained via knowledge distillation so much of that ability can run on modest devices.

The novelty is in balancing accuracy with practicality. Rather than relying on one heavy model, it blends several complementary ones for reliability, then compresses that know-how into a compact model, so strong pest recognition can reach a farmer's phone in the field, not just a data centre.