CNN Filter Refinement for Plant Disease Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current plant disease detection methods using convolutional neural networks (CNNs) face challenges in recognizing multiple diseases across various environmental factors with limited training data, leading to inefficiencies in classification accuracy, especially when symptoms appear in different forms on inoculated and non-inoculated leaves.
Innovation Solution
The system refines a CNN by selecting and ranking filters from the last convolutional layers to focus on class-specific features, constructing vectors for each image, and adding a masking layer to turn off non-essential filters, resulting in a leaner model that can recognize diseases effectively across different environmental conditions without requiring extensive training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a CNN is designed to recognize diseases from inoculated leaves with clean backgrounds, then classification accuracy for inoculated samples is improved, but the model fails to recognize the same diseases from non-inoculated leaves with busy backgrounds
Solution Approach 1:
The patent segments the CNN into two distinct models: a first CNN trained on inoculated leaf images with clean backgrounds, and a second CNN trained on non-inoculated leaf images with busy backgrounds. This segmentation allows each model to specialize in its specific condition, resolving the contradiction by creating multiple specialized classifiers rather than attempting to build a single universal model that would compromise accuracy on either condition.
Solution Approach 2:
The system achieves universality by creating an ensemble of multiple CNN models that can handle different environmental conditions. The first CNN handles inoculated samples while the second CNN handles non-inoculated samples, together forming a universal disease detection system that adapts to various background conditions and sample types without sacrificing accuracy for either category.
2Adaptability or versatility
If multiple CNNs are built to handle different environmental conditions, then adaptability across various factors is improved, but the complexity of the system and computational resources increase
Solution Approach 1:
The patent segments the overall detection task into multiple specialized CNN models, each handling a specific environmental condition. This segmentation improves adaptability by creating specialized models for different conditions while managing complexity through modular design, where each segment can be independently trained and optimized without affecting other segments.
Solution Approach 2:
The system uses an intermediary ensemble mechanism that coordinates multiple specialized CNNs. This intermediary layer manages the complexity by providing a unified interface for disease detection while routing different input types to appropriate specialized models, thereby maintaining adaptability without exposing the full complexity of multiple models to the user.
3Device complexity
If a single CNN is used with limited training data, then device complexity is reduced, but classification accuracy deteriorates when symptoms appear in various forms across different conditions
Solution Approach 1:
The patent segments the training data and model architecture into condition-specific components. By creating separate CNNs for different environmental conditions with dedicated training sets, the system achieves high classification accuracy for each condition without requiring a single massive complex model, thereby improving precision while keeping individual model complexities manageable.
Solution Approach 2:
The system changes parameters by creating multiple models with different architectures and training parameters optimized for specific conditions. This allows each model to achieve high accuracy for its designated condition with appropriate parameter settings, avoiding the need for a single overly complex model while maintaining overall high classification precision across all conditions.
Data Source
AI summary
A system and processing methods for refining a convolutional neural network (CNN) to capture characterizing features of different classes are disclosed. In some embodiments, the system is programmed to start with the filters in one of the last few convolutional layers of the initial CNN, which often correspond to more class-specific features, rank them to hone in on more relevant filters, and update the initial CNN by turning off the less relevant filters in that one convolutional layer. The result is often a more generalized CNN that is rid of certain filters that do not help characterize the classes.