Image-Based Asymmetric Feature Fusion Using EfficientNet-B0 and MobileNetV3Large for Multi-Class Fish Disease Classification
Abstract
This study proposes an asymmetric hybrid feature-fusion network that combines EfficientNet-B0 and MobileNetV3Large in parallel for multi-class classification of eight fish disease categories. The approach is designed to exploit the complementary feature-learning capabilities of both lightweight convolutional neural network architectures while maintaining computational efficiency for potential diagnostic applications. The model was evaluated on a public dataset of 2,400 digital images, divided into training, validation, and test sets using an 80:10:10 ratio. The held-out test set contained 240 unseen images, with 30 images representing each class. As all data partitions originated from the same dataset, the evaluation reflects performance on the same source distribution rather than external generalization. The proposed architecture fuses spatial feature representations through parallel Global Average Pooling (GAP) and Global Max Pooling (GMP) layers. Training followed a progressive two-phase strategy comprising top-layer warm-up and micro deep fine-tuning with a learning rate of 10−5 and label smoothing. Experimental results achieved a validation accuracy of 0.9667 with a loss of 0.5843, and a test accuracy of 0.9833 with a loss of 0.5375 and a macro-average ROC-AUC of 0.9998. On an NVIDIA RTX 3080 GPU, the model recorded an average inference time of 8.61 ms per image, equivalent to 116.18 FPS. Further evaluation using external farm datasets and low-power edge devices is required before real-world deployment.
Related
- HyEff-UNet: an attention-driven Efficientnet U-Net with hybrid loss optimization for UAV-based marine oil spill detection and segmentation
- An explainable swarm-optimized hybrid GRU–Transformer framework with wavelet-based spatial–temporal feature modeling and activation function analysis for rainfall classification across Pakistan’s climatic regimes
- Asymmetric bloom lengthening drives high-latitude ocean chlorophyll accumulation imbalance
- Diurnally asymmetric warming and hydroclimatic stress structure alpine ecosystem vulnerability in the Three-River headwaters region
- Asymmetric basin-wide responses of coastal phytoplankton biomass to the Indian Ocean dipole.
BibTeX
@article{helmud2026imagebased,
title = {{Image-Based Asymmetric Feature Fusion Using EfficientNet-B0 and MobileNetV3Large for Multi-Class Fish Disease Classification}},
author = {Ellya Helmud and Kiswanto},
journal = {Journal of Information Systems and Informatics},
year = {2026},
month = oct,
doi = {10.63158/journalisi.v8i5.1798},
url = {https://doi.org/10.63158/journalisi.v8i5.1798},
note = {CC BY},
}
RIS
TY - JOUR
TI - Image-Based Asymmetric Feature Fusion Using EfficientNet-B0 and MobileNetV3Large for Multi-Class Fish Disease Classification
AU - Ellya Helmud
AU - Kiswanto
PY - 2026
DA - 2026/10/24
JO - Journal of Information Systems and Informatics
DO - 10.63158/journalisi.v8i5.1798
UR - https://doi.org/10.63158/journalisi.v8i5.1798
AB - This study proposes an asymmetric hybrid feature-fusion network that combines EfficientNet-B0 and MobileNetV3Large in parallel for multi-class classification of eight fish disease categories. The approach is designed to exploit the complementary feature-learning capabilities of both lightweight convolutional neural network architectures while maintaining computational efficiency for potential diagnostic applications. The model was evaluated on a public dataset of 2,400 digital images, divided into training, validation, and test sets using an 80:10:10 ratio. The held-out test set contained 240 unseen images, with 30 images representing each class. As all data partitions originated from the same dataset, the evaluation reflects performance on the same source distribution rather than external generalization. The proposed architecture fuses spatial feature representations through parallel Global Average Pooling (GAP) and Global Max Pooling (GMP) layers. Training followed a progressive two-phase strategy comprising top-layer warm-up and micro deep fine-tuning with a learning rate of 10^−5 and label smoothing. Experimental results achieved a validation accuracy of 0.9667 with a loss of 0.5843, and a test accuracy of 0.9833 with a loss of 0.5375 and a macro-average ROC-AUC of 0.9998. On an NVIDIA RTX 3080 GPU, the model recorded an average inference time of 8.61 ms per image, equivalent to 116.18 FPS. Further evaluation using external farm datasets and low-power edge devices is required before real-world deployment.
SN - 2656-5935
SN - 2656-4882
ER -