基于图像增强与GC-YOLO v5s的水下环境河蟹识别轻量化模型研究

Lightweight Model for River Crab Detection Based on Image Enhancement and Improved YOLO v5s

摘要:

利用机器视觉技术识别水下河蟹目标是实现河蟹养殖装备智能化的有效途径之一。针对水下环境目标识别困难、河蟹包含特征信息少、主流的目标检测模型复杂度高等问题，在YOLO v5s的基础上提出了一种适用于水下环境的轻量级河蟹识别模型GC-YOLO v5s（GhostNetV2-CBAM-YOLO v5s）。利用改进的图像增强算法对水下河蟹图像进行预处理以改善其质量；为降低模型复杂度，提出了基于 GhostNetV2的G3模块以改进模型的特征提取网络，并利用幻影卷积进一步轻量化模型；为了优化模型的河蟹特征学习能力，在Neck层和Head层之间引入卷积块注意力模块（Convolution block attention module，CBAM）。实验结果表明，该模型测试集的平均精度均值（Mean average precision，mAP）、召回率和精确率分别为95.61%、97.03%和96.94%，较YOLO v5s分别提升2.80、2.25、2.28个百分点；而GC-YOLO v5s的参数量、浮点运算量和模型内存占用量仅为YOLO v5s的69.1%、56.3%和58.3%。通过实验对比，该模型在识别精度和模型复杂度上优于其他主流目标检测模型；识别速度仅次于YOLO v5s，可达到104f/s。

关键词:水产养殖;河蟹识别模型;图像增强;YOLO v5s;轻量化

Abstract:

Using machine vision technology to identify underwater crab targets is one of the effective ways to achieve intelligent crab farming equipment. However, river crab detection methods face challenges in the difficulty of target detection in underwater environments, limited feature information and high complexity of mainstream target detection models. To solve these challenges, a lightweight river crab detection model GC-YOLO v5s (GhostNetV2-CBAM-YOLO v5s) was proposed. These specific enhancements were as follows: an improved image enhancement algorithm was used to preprocess underwater crab images to improve the detection accuracy;in order to reduce model complexity, a G3 module based on GhostNetV2 was proposed to improve the feature extraction network of the model, and Ghost convolution was used to further lightweight the model;the convolution block attention module (CBAM) was introduced to solve the challenge of extracting deep features within underwater environments, which were integrated into the feature extraction network. The experimental results demonstrated the improved model’s mAP50, recall, and precision on the test set, reaching 95.61%, 97.03% and 96.94%, respectively. These metrics displayed enhancements of 2.80 percentage points, 2.25 percentage points and 2.28 percentage points compared with the baseline. Moreover, GC-YOLO v5s parameters, computations, and model size were only 69.1%, 56.3%, and 58.3% of YOLO v5s respectively. Comparative trials against mainstream object detection algorithms showcased the superiority in accuracy and model complexity. While slightly trailing YOLO v5s in detect speed, GC-YOLO achieved 104f/s.

Key words:aquaculture;river crab detection model;image enhancement;YOLO v5s;lightweight