基于改进 YOLO v8 的泰山赤鳞鱼轻量化计数模型

Lightweight Counting Model for Taishan Red-scaled Fish Based on Improved YOLO v8

摘要:

泰山赤鳞鱼是一种泰山区域特有、具有较高经济价值的高端食用鱼种,在渔获季节和日常养殖中仍以人工计数为主。 为实现养殖和捕捞过程的精准识别计数,提出了一种基于 YOLO v8 的泰山赤鳞鱼轻量化计数模型 ACSS-YOLO v8。 首先,采用 AIFIC 模块加强模型对图像高级特征的处理能力,同时降低模型参数量和计算量。 其次,在 AIFIC 模块后加入 CAFM,通过全局特征和局部特征融合,提升模型的检测性能。 然后,基于 Star Blocks 结构设计新的颈部,增强模型表达能力、降低计算复杂度。 最后,引入 SEAM 提高模型识别遮挡目标的能力,并进一步降低计算冗余和内存消耗。 实验结果表明,本文设计的 ACSS-YOLO v8 模型的精确率、召回率和 mAP50 分别达到了 92. 6% 、85. 8% 和 89. 7% ,相较于 YOLO v8s 模型,其参数量、浮点运算量和内存占用量分别降低了 20% 、25% 和 20% 。 计数模块和 Jetson Orin Nano Super 部署实验表明,可视化的计数模块能够直观显示目标数量,图像检测速度达到 30 f / s,能够满足复杂现场环境的检测需求。

Abstract:

The Taishan red-scaled fish (Varicorhinus macrolepis)is a kind of high-end edible fish species with high economic value unique to Taishan area. It is still mainly counted manually in the fishing season and daily breeding. In order to realize the accurate identification and counting of the breeding and fishing process, a lightweight identification model ACSS-YOLO v8 was proposed for Taishan red-scaled fish based on YOLO v8. Specifically, multiple structural optimizations were integrated to balance detection accuracy and computational efficiency simultaneously. Firstly, the AIFIC module was used to enhance the model's ability to process high-level image features, while reducing parameter volume and calculations. Secondly, CAFM was added after the AIFIC module. By fusing global and local features, the detection performance of the model was enhanced. Then a neck network was designed based on the star blocks structure. This design enhanced the feature expression ability of the model and further reduced the computational complexity. Finally, SEAM was introduced to improve the model's ability to recognize occluded objects, and further reduce computational redundancy and memory consumption. The experimental results showed that the P, R and mAP50 of the ACSS-YOLO v8 model designed reached 92. 6% , 85. 8% and 89. 7% , respectively. Compared with the YOLO v8s model, the number of parameters, GFLOPs and memory usage were reduced by 20% , 25% and 20% , respectively. The counting module was developed to provide intuitive visual feedback for practical application scenarios. The experiments on the counting module and Jetson Orin Nano Super deployment showed that the visualized counting module can intuitively display the number of objects, and the image inference speed reached 30 f / s, which can meet the detection requirements of complex on-site environments.