An optimized domain-specific shrimp detection architecture integrating conditional GAN and weighted ensemble...

Area of Science:

• Computer Vision
• Machine Learning
• Aquaculture Technology

Background:

• Deep learning excels at image pattern recognition using pixel intensities.
• Improving object detection accuracy often requires parameter tuning or synthetic data generation.
• Existing methods face challenges in accurately detecting shrimps within diverse image backgrounds.

Purpose of the Study:

• To introduce an "enhanced shrimp detection using integrated augmentation (ESDIA)" approach for improved shrimp detection.
• To leverage synthetic data generation to augment datasets and enhance model robustness.
• To combine segmentation techniques with advanced deep learning classifiers for precise shrimp identification.

Main Methods:

• Image segmentation techniques including grayscale conversion, edge detection, and morphological operations.
• Dataset construction incorporating diverse shrimp images and backgrounds.
• Implementation of deep learning classifiers such as Faster Recurrent Convolution Neural Network (FRCNN) and YOLOv7.
• Synthetic data generation using Generative Adversarial Networks (GANs) to increase dataset variance.

Main Results:

• The ESDIA approach demonstrated a vital flow in object detection rates.
• Achieved a mean average precision (mAP) ranging from 80.53% to 89.13%.
• Validated the efficacy of integrated augmentation and deep learning in elevating shrimp detection.

Conclusions:

• The proposed ESDIA method effectively enhances shrimp detection capabilities.
• Synthetic data generation significantly bolsters dataset volume and variance, improving model performance.
• The research highlights the potential of deep learning paradigms in advanced aquaculture applications.