JassidNet

The repository is for the papaer: JassidNet: A Quantization-Aware Lightweight Phenotyping Framework for High-Throughput Cotton Jassid (Amrasca biguttula) Detection and Counting Toward Objective Resistance Screening, including the code and a benchmark subset for reproducing.

This paper is under review on Industrial Crops and Products.

Problem Statement

Fig.1 Illustration of background complexity in field-acquired cotton leaf images and the target-specific focus of the proposed approach.

Dataset and Dataset Construction

We propose a pseudo-label iteration–based dataset construction strategy and introduce the first vision dataset dedicated to cotton jassid recognition, termed the Cotton Jassid Recognition (CJR) Dataset.

The dataset will be publicly released on [Kaggle](coming soon).

Fig.2 Workflow of semi-supervised dataset construction with iteration-based pseudo-labels. Note: Flows in the loop are highlighted by pink arrows; pink dashed line indicates that the operation is performed only once; the yellow star marks the start of the entire pipeline.

Fig.3 Representative images of CJR dataset samples. Note: Red circles represents adults (longer length, presence of wings and the characteristic two black spots), and yellow represents nymphs (smaller size when compared to adults and absence of wings).

JassidNet Pipeline Overview

Fig.4 Workflow overview of JessidNet. Note: Detector-guided/Text-Driven zero-shot segmentation function is highlighted in green dash box, pink dashed arrow indicates optional super-resolution enhancement function, purple dashed arrow indicates the branch of Detector-Guided Zero-Shot Segmentation function, blue dashed arrow indicates the optional branch of Text-Driven Zero-Shot Segmentation function.

Fig.5 Structural diagram of proposed Smart Jassid Inference and Monitoring Network (SJIMNet) and its key modules. Note: modified modules are highlighted by black wireframe.

Explainability and Qualitative Results

Fig.6 Feature response visualization of baseline model and the proposed SJIMNet-O under representative jassid densities scenarios: (a) low-density; (b) medium-density; (c) high-density.

Fig.7 Representative visualization examples of JassidNet on External Test Set.

Biological and Breeding Applications

Fig.8 Weekly boxplots of jassid counts per leaf across five cotton genotypes in the first external test set (October 2025).

Pretrained Weights

The pretrained weights of SJIMNet-O (FP32) and SJIMNet (INT8) are provided in the weights/ directory:

• SJIMNet-O.pt: Optimized model trained on the Cotton Jassid Recognition (CJR) dataset.
• SJIMNet.pth: Quantized customized model trained on the CJR dataset.

These weights are intended for inference, visualization (e.g., Grad-CAM), and downstream biological analysis.

Getting Start

See [JassidNet Handbook: A Practical Guide for Field Data Collection, Detection, and Phenotyping of Cotton Jassids (Amrasca biguttula)](coming soon).

Citation

JassidNet Paper: coming soon.....
JassidNet Handbook: coming soon....