OMSACC-SGRAN: an implementation of hybrid Optimized Multi-Scale Atrous Convoluted CNN with Self Guided Residual Attention Network for fish species classification

Visualizing the behavior of fish variants is the primary significance for obtaining biological insights in the marine ecological system. Several computer vision and machine learning-based approaches are introduced to classify the fish variants, but these require large data sets to provide high classification accuracy. Deep structured models handle these issues, but they need more attention during training time because some useful information is missed during the training process. It makes it possible to classify the fish variants inaccurately. Therefore, an automatic fish species classification system is developed to provide high accuracy and avoid misclassification results. Thus, this research work explores a new fish species classification strategy with the adoption of a hybrid deep learning technique. The acquired images are pre-processed using contrast-limited Adaptive Histogram Equalization (CLAHE) and histogram equalization methods for cleaning and increasing the quality of images. These images are fed to the hybrid classifier known as Optimized Multi-Scale Atrous Convoluted CNN with Self-Guided Residual Attention Network (OMSACC-SGRAN). Here, the parameters of both Multi-scale Atrous Convoluted CNN and Multi-scale Self Guided Residual Attention Network are tuned using a newly recommended Marine Predators-Forensics-based Investigation inspired Algorithm (MP-FBI). The experimental results proved that this approach achieve excellent results.