A Hybrid Robust Learning Framework for Multi-Scale Chicken Detection in Extreme Battery Cage Environments Using Augmented Transfer Learning
Abstrak
Multi-scale chicken detection in battery-cage farms is a challenging computer vision problem due to the severe occlusion from cage bars, low and inconsistent illumination, and high object density. This study proposes a hybrid robust learning framework based on augmented transfer learning to improve multi-scale chicken detection. The detector is built on Faster Region-based Convolutional Neural Networks (R-CNN) with a ResNet-50–FPN backbone. It is enhanced by an early-dilated convolutional block and a regularized custom head incorporating dropout and batch normalization. Transfer learning is applied through selective fine-tuning, in which only backbone stages 3–4 are retrained while earlier layers are frozen to preserve general representations. Four training schemes were evaluated: (a) baseline without augmentation, (b) online dynamic augmentation, (c) offline augmentation using Albumentations, and (d) hybrid offline+online augmentation. The augmentation pipeline includes rotation, horizontal flip, brightness/contrast adjustment, Gaussian blur, gamma correction, and resizing to 512×512, with COCO-format bounding boxes synchronized during transformation. Experiments using COCO evaluation show that the hybrid scheme achieves the best overall in-domain performance, reaching mAP@0.5 = 0.880, mAP@[0.50:0.95] = 0.414, and AR@100 = 0.483, corresponding to an absolute gain of 0.098 in mAP@0.5 over the offline-only scheme. Despite these gains, small-object detection remains unresolved, indicating the need for higher-resolution feature representations and anchor/grid redesign for micro-scale and heavily occluded instances. Overall, the findings indicate improved detection performance and in-domain generalization under naturally challenging battery-cage conditions. However, no dedicated corruption-based robustness benchmark or stress test was conducted; therefore, the results should not be interpreted as formal evidence of robustness to explicit illumination, blur, or noise perturbations.
Identifier
- OpenAlex: https://openalex.org/W7163760699
- DOI: https://doi.org/10.48084/etasr.17563