基于多尺度融合的卷积神经网络的杂草幼苗识别

Weed Seeding Recognition Based on Multi-Scale Fusion Convolutional Neutral Network

摘要: 针对传统网络ALexNet识别精度不高、内存需求量大、特征尺度单一等问题，该文提出了一种多尺度并行融合的轻量级模块PL-SE模块。该模块将上层特征经过两个不同尺度的卷积核和一个最大池化，融合之后得到新的特征信息，之后再经过一个SE模块，最后进行残差学习。同时，对于下采样部分，该文提出一种SR-SE模块代替传统网络的池化层，在降维的同时进行特征提取。使用PL-SE模块和SR-SE模块对ALexNet模型改进得到一种新的模型，用于对25种杂草幼苗进行训练识别。改进后的模型识别准确率达到了96.32%，相较于传统的ALexNet模型提高了8个百分点，参数总量减少约56.7 M (Million)。除此之外，与ResNet、GoogleNet、MobileNet等经典网络相比，改进后的模型在准确率和参数量方面都具有优势。

Abstract: For the problems that the traditional network AlexNet recognition accuracy is not high, memory demand is large, the characteristic scale is single and so on, this paper presents a multi-scale parallel fusion lightweight module PL-SE module. In this module, the upper features are processed by two convolution kernels with different scales and a maximum pooling. After fusion, new feature information is obtained, followed by an SE module, and finally residual learning is conducted. At the same time, for the lower sampling part, this paper proposes a SR-SE module instead of the pooling layer of traditional network, which can extract features while reducing dimensions. Using PL-SE module and SR-SE module to improve ALexNet model, a new model was developed for training and identification of 25 weed seedlings. The improved model recognition accuracy reaches 96.32%, 8 percentage points higher than the traditional ALexNet model, and the total number of parameters is reduced by about 56.7 m (Million). In addition, compared with classic networks such as ResNet, GoogleNet and MobileNet, the improved model has advantages in accuracy and number of parameters.