<p>In marine life, images classification is an important task for environmental monitoring and marine resource management; it has significant implications for real-time monitoring, tracking endangered species, and ensuring the effectiveness of conservation efforts. However, the accurate classification is difficult due to challenges like light distortion, poor visibility, and limited annotated data. To improve classification accuracy, we introduced a dual-path feature fusion neural network architecture inspired by squeeze and excitation operations. Our model employs a dual-path approach for feature extraction: the first path utilizes Xception model by initially freezing to keep weights and fine-tune for new datasets. The second path used a custom CNN that implements multiple convolutions, pooling, and Squeeze-and-Excitation operations to abstract complementary features. The outputs of both paths are concatenated as feature vector. The classification stage involves fully connected layers with activation functions and regularization, applying dropout to prevent overfitting, and culminating in a softmax layer for final classification. Experiments conducted on the LifeCLEF2015 and Fish4Knowledge datasets demonstrate competitive results, with accuracies of 99.57% and 99.39%, respectively, and more than 99% overall recall, precision, and F1 score. The model generalized well with other general datasets and achieved high accuracy. Different visualization techniques are used to validate the effectiveness of our approach. Although the model shows promising results, challenges remain, including the need for high-quality labeled data and the model’s performance in extremely low visibility conditions. The future research could focus on enhancing data augmentation and domain adaptation strategies. Additionally, tuning the hyperparameters proved to be quite challenging.</p>