朱道恒, 朱浛灵, 李志强, 刘润. 基于改进YOLOv7的海滩垃圾检测方法[J]. 海洋环境科学, 2024, 43(5): 684-695, 706. DOI: 10.12111/j.mes.2023-x-0224
引用本文: 朱道恒, 朱浛灵, 李志强, 刘润. 基于改进YOLOv7的海滩垃圾检测方法[J]. 海洋环境科学, 2024, 43(5): 684-695, 706. DOI: 10.12111/j.mes.2023-x-0224
ZHU Daoheng, ZHU Hanling, LI Zhiqiang, LIU Run. A beach litter detection method based on improved YOLOv7[J]. Chinese Journal of MARINE ENVIRONMENTAL SCIENCE, 2024, 43(5): 684-695, 706. DOI: 10.12111/j.mes.2023-x-0224
Citation: ZHU Daoheng, ZHU Hanling, LI Zhiqiang, LIU Run. A beach litter detection method based on improved YOLOv7[J]. Chinese Journal of MARINE ENVIRONMENTAL SCIENCE, 2024, 43(5): 684-695, 706. DOI: 10.12111/j.mes.2023-x-0224

基于改进YOLOv7的海滩垃圾检测方法

A beach litter detection method based on improved YOLOv7

  • 摘要: 海滩垃圾污染对海洋生态系统和人类健康构成巨大威胁,海滩垃圾监测和清理是一项繁重且复杂的任务。传统的人工调查存在监测效率低、检测范围小和时效性差等问题,为了解决这些问题,本文提出一种基于改进YOLOv7的海滩垃圾检测方法。首先,YOLOv7与量化感知RepVGG(QARepVGG)相结合,实现快速计算并降低模型参数量。其次,加入简单注意力机制(simple attention mechanism,SimAM),增强网络对图像感兴趣区域的特征提取能力。最后,结合双向特征金字塔网络(bi-directional feature pyramid network, BiFPN)结构,改进原有路径聚合网络(path aggregation network, PAN),提高网络学习垃圾特征的效率,增强对不同尺寸垃圾的识别能力。在自建数据集上的实验结果表明:(1)改进模型对8类海滩垃圾有良好的检测能力;(2)与YOLOv7相比,改进模型的总体平均精度均值(mean average precision, mAP)提升5.8%,每秒传输帧数(frames per second, FPS)提高17,改进模型对泡沫、塑料类和纸制品垃圾的识别精度最高;(3)与几种流行的检测模型相比,改进模型的识别精度和效率最高。实际场景中的检测结果表明,改进模型能满足海滩垃圾实时性检测需求。

     

    Abstract: Beach litter monitoring and cleaning is a difficult and involved undertaking since beach litter contamination poses a serious hazard to human health and marine ecosystems. Due to the low monitoring efficiency, limited detection range, and unreliable timeliness of typical human surveys, this work proposes a beach litter detection method based on improved YOLOv7. First, YOLOv7 is initially merged with Quantitative awareness RepVGG (QARepVGG) to achieve rapid computation and reduce the number of model parameters. Second, the SimAM attention mechanism is included to improve the network's capacity for feature extraction from the targeted region of the picture. To further enhance the original path aggregation network (PAN), the Bi-directional feature pyramid network (BiFPN) structure is merged with it. As a result, the network is more effectively able to identify targets of various sizes and recognize the characteristics of different litter sizes. Experimental findings on the self-constructed dataset demonstrate that the modified model has strong identification capability for eight different categories of beach litter. The revised model outperforms YOLOv7 in terms of mean average precision (mAP) improvement by 5.8%, frame per second (FPS) improvement by 17, and the improved model has the highest recognition accuracy for styrofoam, plastic, and paper. When compared to a number of widely used detection models, our improved model has the best performance. These detection results in real scenarios show that the improved model is capable of detecting beach litter in real time.

     

/

返回文章
返回