基于多尺度特征融合和注意力机制的水面死鱼检测方法

A Method for Detecting Dead Fish on Water Surfaces Based on Multi-scale Feature Fusion and Attention Mechanism

死鱼对于水域生态和饮水安全存在巨大威胁,由于水面环境复杂,导致现有目标检测算法在死鱼检测任务中存在漏检、误检等情况。为此,以单次多边框检测(single shot multibox detector,SSD)为基础提出一种基于多尺度特征融合和注意力机制的水面死鱼检测方法FFA-SSD(SSD with feature fusion and attention)。首先,采用计算量和参数量更少且特征提取能力更强的残差网络ResNet50替换VGG16主干网络;其次,设计了多尺度特征融合模块,增强浅层特征和高层语义信息的融合;然后,引入通道注意力机制,抑制特征融合带来的冗余信息干扰,提升网络对目标的关注度;最后,设计了一种适用于小目标检测的数据增强算法,扩充训练数据中的小目标数量,丰富训练背景。实验结果表明,同现有目标检测算法相比,FFA-SSD算法可以更好地识别水面死鱼,检测精度达到93.5%。

Dead fish could pose a huge threat to water ecology and safety of drinking water.With complex water surface environment,the existing object detection algorithms had some flaws such as missed and false detections in small target.Therefore,a dead fish detection method on water surface based on the multi-scale feature fusion and attention mechanism,SSD with feature fusion and attention(FFASSD)was proposed.Firstly,the residual network ResNet50 with less computation and fewer parameters and better feature extraction ability was used to replace the VGG16 backbone network.Then,a multiscale feature fusion module was designed to enhance the fusion of shallow features and high-level semantic information.Finally,a channel attention mechanism was introduced to suppress the interference of redundant information brought by feature fusion and to improve the network′s focus on the target.In addition,a data enhancement algorithm applicable for small target detection was designed to increase the number of small targets in the training data and to enrich the training background.The experimental results showed that compared with other target detection algorithms,the recognition function of FFA-SSD algorithm for dead fish on the water surface was better,and the detection accuracy was at 93.5%.