A CenterNet2-Based Approach for Multi-Class Chili Ripeness Detection in Real Agricultural Settings

Abstract

This study develops a computer-vision–based system for detecting and classifying the ripeness levels of chili peppers (Capsicum annuum L.) using the CenterNet2 architecture with three backbone/decoder configurations: ResNet50–FPN, DLA–BiFPN, and Res2Net101–DCN. These configurations were compared to assess the influence of multi-scale feature structures and the integration of Deformable Convolution Networks (DCN) on the model’s capability to detect small fruits and handle leaf occlusion. The objective of this research is to improve the accuracy of chili fruit detection and ripeness classification in agricultural environments characterized by complex backgrounds, inconsistent lighting, and partial obstruction by leaves or stems. The dataset consists of 659 field images annotated into three ripeness categories: ripe, almost ripe, and unripe. Model performance was evaluated using standard COCO metrics, including Average Precision (AP), AP50, and AP75. Experimental results show that the CenterNet2–Res2Net101–DCN configuration achieves the highest performance, obtaining an AP of 66.6%, AP50 of 90.3%, and AP75 of 81.7%. It also attained category-specific classification accuracies of 63.2% for ripe, 66.1% for almost ripe, and 70.5% for unripe chili peppers. These findings demonstrate that incorporating DCN into the CenterNet2 framework significantly enhances the detection of small objects and improves visual classification under challenging field conditions. The outcomes of this study are expected to support the development of AI-based automated harvesting systems and precision agriculture applications in the future.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint