A multi-scale gated attention and wavelet transform-based lightweight network for efficient plankton classification

Abstract

Plankton monitoring is vital for ocean resource exploration. However, plankton images are numerous and often lack detail, making traditional classification methods prone to low accuracy and inefficiency, while ensemble learning struggles in resource-constrained environments. This study proposes a lightweight plankton image classification method based on GhostNet, named MGLR-WT-CBF-GhostNet (MWC-GhostNet), to enhance classification efficiency and reduce computational overhead. The method introduces the Multi-Scale Gated Long-Range (MGLR) module, which aggregates long-range dependencies using multi-scale asymmetric convolution kernels and learns feature weights via gating mechanisms. This fusion of multi-scale information enhances the model's ability to capture key image features. Additionally, wavelet transform is applied to depthwise convolution in the expansion layer to extract multi-frequency information, improving the model's ability to capture plankton shapes and contours. Class-Balanced Focal Loss is employed to address class imbalance and enhance recognition of hard samples. Experimental results show that MWC-GhostNet achieved an accuracy of 78.43 % and an F1 score of 0.6702 on the Kaggle121 dataset, 98.81 % accuracy and 0.9618 F1 score on the PMID2019 dataset, and 89.13 % accuracy and 0.9128 F1 score on the ZOOSCAN20 dataset. Compared to the baseline models, the proposed method improves accuracy by 1.81 %, 3.03 %, and 3.49 %, and enhances the F1 score by 5.83 %, 4.93 %, and 5.60 %, respectively. The model demonstrates high efficiency and is well-suited for deployment in resource-constrained environments.