基于深度残差网络的水面漂浮物检测

Detection of Floating Objects on Water Surface Based on Depth Re-Sidual Network

摘要: 水面漂浮物作为水体污染的重要来源，严重威胁着水域生态安全和人类健康。针对现有检测方法在复杂水面场景下存在特征提取不足和小样本适应性弱的问题，本研究提出了一种改进的深度残差网络检测框架。首先构建自主采集的多源水面漂浮物数据集，通过自适应图像裁剪和归一化处理消除光照波动与尺度差异的干扰；继而基于ResNet-18架构设计分层优化策略，冻结浅层网络参数保留通用特征表征能力，通过微调深层网络增强域适应特征提取，并构建复合分类头提升小样本学习性能。针对数据集规模限制，创新性融合多维度协调数据增强策略与类别平衡损失函数，通过在线增强扩充数据多样性并缓解类别不均衡问题。实验表明，该方法在自建数据集上取得98.71%的检测准确率，较基线模型提升9.68个百分点，且在不同光照条件和漂浮物密度的测试场景中均保持稳定性能。研究系统验证了从数据预处理、网络架构优化到训练策略设计的全流程方案有效性，提出的迁移学习框架在保证模型轻量化的同时显著提升检测鲁棒性，为构建智能化水面污染监测系统提供了可靠的技术方案。

Abstract: As a significant source of water pollution, floating objects on water surfaces pose severe threats to aquatic ecological security and human health. To address the limitations of existing detection methods, such as insufficient feature extraction in complex aquatic scenarios and weak adaptability to small-sample conditions, this study proposes an improved deep residual network-based detection framework. Firstly, a self-collected multi-source dataset of floating objects was constructed, where adaptive image cropping and normalization processing were applied to mitigate interference from illumination fluctuations and scale variations. Subsequently, based on the ResNet-18 architecture, a hierarchical optimization strategy was designed: shallow network parameters were frozen to retain general feature representation capabilities, while deep networks were fine-tuned to enhance domain-adaptive feature extraction, coupled with a composite classification head to improve small-sample learning performance. To overcome dataset size constraints, a novel integration of multi-dimensional coordinated data augmentation strategies and class-balanced loss functions was implemented, expanding data diversity through online augmentation and alleviating class imbalance issues. Experimental results demonstrate that the proposed method achieves a detection accuracy of 98.71% on the self-built dataset, representing an improvement of 9.68 percentage points over the baseline model, while maintaining stable performance across varying illumination conditions and floating object densities. The study systematically validates the effectiveness of the full-process solution encompassing data preprocessing, network architecture optimization, and training strategy design. The proposed transfer learning framework significantly enhances detection robustness while ensuring model lightweight design, providing a reliable technical solution for intelligent water pollution monitoring systems.