融合轻量级网络的农业病害检测

Agricultural disease detection integrated into lightweight networks

摘要:

农业病害会导致作物脱叶早,光合作用减弱,从而影响作物质量,减少农民收入。针对病害初发期间目标小、背景复杂和室外光线变化大导致的目标误检问题,本文提出一种融合轻量级网络的YOLOv4检测算法。首先对主干网络进行剪枝并增加多尺度的分组卷积提高模型对复杂背景的抗干扰性,其次设计轻量级SCE(space channel expand) 注意力机制降低深层网络中细节信息丢失的影响。最后设计跳跃连接特征金字塔(jump connection feature pyramid network,JC-FPN) 替换PAnet(path aggregation network)特征融合模块从而进一步实现模型轻量化。实验结果表明,改进算法在本文数据集上的mAP₅₀达到了84.17%,检测速度为50 FPS,相比于YOLOv4检测算法分别提高了0.71%和10 FPS,满足移动端对农业病害的检测精度和速度的要求。

Abstract:

Agricultural diseases can cause early defoliation of crops and weakened photosynthesis,which can affect crop quality and reduce farmers′ incomes.Aiming at the problem of target miss detection exception caused by small targets, complex background and unstable natural light during the initial occurrence of the diseases,this paper proposes a YOLOv4 detection algorithm that integrating lightweight networks.Firstly,the trunk network is simplified and multi-scale group convolution is enhanced to improve the anti-interference ability of the mode in the complex backgrounds.Secondly,the lightweight space channel expand (SCE) attention mechanism is designed to reduce the impact of detail information loss in the deep network.Finally,the pyramid structure with the feature of skip connection is applied for the replacement of integration module with path aggregation network (PAnet) feature to further realize the model lightweight.Experimental results show that the improved algorithm reaches 84.17% of mAP₅₀ and the detection speed is 50 FPS in the dataset of this paper,which is 0.71% and 10 FPS higher than that of YOLOv4 detection algorithm,that meets the requirements of the detection accuracy and speed of agricultural diseases on the mobile devices.