【摘要】 针对现有害虫识别模型复杂度高、计算量和参数量巨大的问题，提出一种基于DCP-ShuffleNetV2的轻量级森林害虫识别模型。该模型主要从特征提取、特征融合、轻量化方面进行改进。首先通过引入金字塔分割注意力模块PSA提取多尺度的空间信息和跨通道依赖关系，有效地学习上下文信息；其次将基准网络模型ShuffleNetV2的Stage模块修改为CSP结构，增强特征融合能力；将模型的普通卷积替换为动态卷积，压缩模型参数量和计算量。试验以雄安新区“千年秀林”害虫为研究对象，构建30类常见害虫数据集。结果表明，改进后的DCP-ShuffleNetV2模型在自制的Forest30数据集上的害虫识别准确率是92.43%,模型参数量、计算量和内存大小分别是0.13 M、24.53 M和9.53 MB,相比于基准网络模型，识别准确率提升3.11%,参数量、计算量和内存大小分别减少62.83%、42.48%和15.13%。与目前常用的分类模型相比，识别准确率平均提高5.39%,模型参数量、计算量和内存大小平均减小14.32 M、1 035.80 M和35.98 MB。更多还原

【Abstract】 This paper proposes a pest recognition model based on lightweight DCP-ShuffleNetV2 to solve the problems of high complexity of model, large amount of computation and reference. The model is improved from feature extraction, feature fusion and lightweight. Firstly, to learn context information effectively, the Pyramid Split Attention(PSA) module is introduced to extract multi-scale spatial information and cross-channel dependency. Secondly, to enhance the feature fusion capability, the Stage module of the benchmark network model ShuffleNetV2 is modified to Cross Stage Partial(CSP) structure. Finally, to compress the number of parameters and computation, the regular convolution is replaced by dynamic convolution for the model. In the experiment, a data set of Forest 30 was constructed in the “Millennium Xiulin” of Xiong’an New Area. The experimental results show that the pest identification accuracy of the DCP-ShufflenetV2 model is 92.43%, and the number of parameters, computation amount and memory size of the improved model are 0.13 M, 24.53 M and 9.53 MB, respectively. Compared with the ShufflenetV2 network model, the pest identification accuracy of the improved model increased by 3.11%, and the reference number, computation amount and memory size were reduced by 62.83%, 42.48% and 15.13%, respectively. Compared with the current commonly used classification model, the average recognition accuracy is increased by 5.39%, the number of parameters, computation amount and memory size of the improved model are reduced by 14.32 M, 1 035.80 M and 35.98 MB on average.更多还原