Classification of microplanktons in an imbalanced digital holographic image dataset with a deep network using channel...

Classifying microplanktons in digital holographic images is challenging due to a multitude of factors. For instance, shifts in viewpoint can alter how microplanktons are perceived, while illumination changes can affect the visibility of certain features. Geometric anomalies can distort the shapes of these microplanktons, and the presence of noise within the digital holographic microscope can further alter local image features. Additionally, the difficulty in data collection results in dataset imbalance leading to a biased classification problem. These class-imbalanced datasets pose a considerable hurdle in machine learning applications. Here, categorical representations tend to favor majority classes while neglecting minority classes that are equally important for a comprehensive understanding of microplankton diversity. Accordingly, this research contributes what we believe to be a novel debiasing method using channel attention blocks (DCABs) and a novel attention product. It enhances the model's ability to focus on relevant features while mitigating the effects of bias. This method was applied on six biased models, viz., VGG16, ResNet50V2, ResNet152V2, InceptionV3, Xception, ShuffleNetV2, and ShincNet. The proposed method achieved a significant reduction in the degree of bias (DoB) and KL divergence (KL) values for all the six biased models. With just 6.68M parameters and 6.4 GFLOPs, the DCAB for ShincNet demonstrated a competitive performance in terms of DoB (0.125) and KL (0.82) compared to four state-of-the-art debiasing techniques.