Sea Cucumber Detection Algorithm Based on Deep Learning.

Area of Science:

• Computer Vision
• Marine Biology
• Machine Learning

Background:

• Traditional Single Shot MultiBox Detector (SSD) for sea cucumber recognition suffers from limited feature expression, high computational load, and challenges in embedded system deployment.
• Existing methods struggle with detailed feature extraction and accurate localization, particularly for smaller targets.

Purpose of the Study:

• To develop an improved sea cucumber detection algorithm addressing the limitations of traditional SSD.
• To enhance feature representation and computational efficiency for real-time embedded applications.

Main Methods:

• An improved SSD algorithm utilizing MobileNetv1 as the backbone.
• Integration of Receptive Field Blocks (RFB) to expand feature receptive fields and capture small target details.
• Incorporation of an attention mechanism to strengthen salient features and suppress irrelevant ones across different network depths.

Main Results:

• The improved algorithm demonstrated superior performance over the traditional SSD, with a 5.1% increase in average precision.
• Enhanced robustness and improved performance on the Precision-Recall (P-R) curve compared to YOLOv4 and Faster R-CNN.
• The algorithm achieves stable, real-time sea cucumber detection with reliable feedback.

Conclusions:

• The proposed enhanced SSD algorithm effectively overcomes the limitations of traditional methods for sea cucumber detection.
• The integration of RFB and attention mechanisms significantly boosts detection accuracy and robustness.
• This improved algorithm is suitable for real-time, embedded sea cucumber monitoring systems.