FE-YOLO: An Efficient Deep Learning Model Based on Feature-Enhanced YOLOv7 for Microalgae Identification and...

Area of Science:

• Marine Biology
• Biotechnology
• Computer Vision

Background:

• Accurate microalgae identification and detection are crucial for harnessing microalgae resources.
• Traditional methods face significant limitations in efficiency and accuracy.
• A lack of comprehensive microalgae cell datasets hinders research and application.

Purpose of the Study:

• To develop an advanced deep learning model for microalgae cell identification and detection.
• To enhance the feature extraction capabilities and convergence stability of existing object detection models.
• To create a valuable dataset for training and evaluating microalgae detection algorithms.

Main Methods:

• Integration of the Coordinate Attention Group Shuffle Convolution (CAGS) module into the YOLOv7 Neck for improved feature extraction.
• Implementation of the SCYLLA-IoU (SIoU) loss function to ensure stable model convergence.
• Construction of a novel microalgae dataset comprising 6300 images across seven species.

Main Results:

• The Feature-Enhanced YOLOv7 (FE-YOLO) model demonstrated significant improvements over the standard YOLOv7.
• Achieved increases of 9.6% in average Precision, 1.9% in Recall, 9.7% in mAP@50, and 6.9% in mAP@95.
• Reduced the average detection time per image by 9.2% to 0.0455 seconds.

Conclusions:

• The proposed FE-YOLO model offers a superior solution for microalgae identification and detection.
• The enhanced model provides higher accuracy and efficiency compared to conventional methods.
• This work contributes a valuable dataset and a robust model for advancing microalgae research and applications.