CNN Mosquito Classification Using Anatomical Landmark Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for classifying insects, particularly mosquitoes, using visual attributes face challenges due to their slender bodies and long legs, leading to difficulties in discrimination and classification, especially when using classical object detection algorithms and machine learning techniques that rely on heuristics and struggle with overlapping insects.

Innovation Solution

A computer-implemented method utilizing a convolutional neural network (CNN) predictive model to localize, classify, and count mosquitoes by training on labeled image patches, allowing for the identification and removal of female mosquitoes from a group before sterilizing male mosquitoes, using image patches and a heuristic filter to improve accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If classical object detection algorithms are used to classify insects, then the classification process can be automated, but the accuracy is insufficient for insects with slender bodies and long legs like mosquitoes

Engineering Contradiction:
Improveautomation of classificationVSAvoidclassification accuracy
Core Design Contradiction:
Extent of automationVSMeasurement precision

Solution Approach 1:

The patent transforms the classification problem from detecting entire insect bodies to detecting specific anatomical landmarks (head and tail endpoints). This parameter change in what is being detected allows accurate classification of mosquitoes despite their slender bodies and long legs, resolving the contradiction between automation and accuracy.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent segments the insect detection task into identifying specific anatomical features (head and tail) rather than detecting the whole insect. This segmentation approach enables accurate classification of overlapping insects with challenging morphologies by focusing on discrete, identifiable landmarks.

Inventive Principle:
Principle #1Segmentation

2Extent of automation

If machine learning techniques relying on heuristics are used, then classification can be performed, but the system cannot reliably localize and classify overlapping insects

Engineering Contradiction:
Improveclassification capabilityVSAvoidreliability with overlapping insects
Core Design Contradiction:
Extent of automationVSReliability

Solution Approach 1:

The patent changes the detection parameters from whole-insect bounding boxes to specific anatomical landmark coordinates (head and tail endpoints). This parameter transformation enables reliable localization and classification of overlapping insects by focusing on discrete, separable features rather than entire insect silhouettes.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent moves from 2D bounding box detection to detecting specific points (endpoints of alignment axis) along the insect body. This dimensional approach to landmark detection provides more precise localization information that remains reliable even when insects overlap in the image plane.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Measurement precision

If visual inspection by humans is used, then classification accuracy can be maintained, but the process is time-consuming and less efficient

Engineering Contradiction:
Improveclassification accuracyVSAvoidclassification speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent transforms the visual inspection task from holistic insect assessment to detecting specific anatomical landmarks (head and tail endpoints). This parameter transformation enables automated systems to achieve human-level accuracy by focusing on discrete, easily detectable features that can be processed quickly by algorithms.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent uses image patches as digital copies of insect regions containing anatomical landmarks. These patches can be processed automatically by computational algorithms, replacing manual visual inspection while maintaining the precision of human experts through automated landmark detection.

Inventive Principle:
Principle #26Copying

4Extent of automation

If training data developed using heuristics is used, then machine learning models can be trained, but the models fail to generalize to all insect populations

Engineering Contradiction:
Improveautomated classificationVSAvoidapplicability to different insect populations
Core Design Contradiction:
Extent of automationVSAdaptability or versatility

Solution Approach 1:

The patent changes the training objective from detecting entire insects to detecting anatomical landmarks (head and tail endpoints with alignment axes). This parameter change creates a more robust representation that generalizes across different insect populations, as anatomical landmarks are universally present and detectable regardless of species-specific variations in body shape or size.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3532962B1Predictive models for visually classifying insects
Publication Date: 2023.06.28 VERILY LIFE SCIENCES LLC
  • EP3532962B1 patent drawingFigure 1
  • EP3532962B1 patent drawingFigure 2
  • EP3532962B1 patent drawingFigure 3

AI summary

Insects can be localized and classified using a predictive model. To begin, image data is obtained that corresponds to the insects.. Using a predictive model, samples of the image data are evaluated to determine whether the image portions include an insect and, if so, into what category the insect should be classified (e.g., male/female, species A/species B, etc.).