This work presents the development and evaluation of two custom convolutional neural network models for detecting the marine species Starfish, based on pre-trained YOLO architectures. The first model was implemented using Yolov4-Tiny in Python and trained on Google Colab with GPU acceleration, while the second was developed in MATLAB using the YOLOv2 framework. A dataset of 500 labeled images was used for training, with annotations performed via LabelImg (Python) and Image Labeler (MATLAB), ensuring proper object recognition. Experimental results demonstrate that the Yolov4-Tiny model achieved superior performance, reaching a detection accuracy of 80.40% in underwater tests, with a confusion matrix accuracy of 90.66%, precision of 93.79%, and sensitivity of 96.45%. In contrast, the MATLAB-based model achieved only 58.63% detection accuracy underwater, with confusion matrix results of 80% accuracy, 87.59% precision, and 90.22% sensitivity. These findings indicate a clear advantage in using Yolov4-Tiny for real-time embedded detection tasks, particularly in resource-constrained environments like underwater robotics. Consequently, the Yolov4-Tiny model was selected for deployment in an autonomous Blue-ROV underwater robot, enabling real-time artificial vision for Starfish identification. The implementation contributes to marine species monitoring and robotic autonomy in challenging underwater conditions.