基于双池化与多尺度核特征加权CNN的典型牧草识别

针对自然背景下牧草难识别的问题,提出一种基于双池化与多尺度核特征加权的卷积神经网络牧草识别方法。双池化特征加权结构通过将卷积层输出的特征图分别进行最大值池化和均值池化得到两组特征图,引入特征重标定策略,依照各通道特征图对当前任务的重要程度进行加权,以增强有用特征、抑制无用特征;多尺度核特征加权结构通过在卷积层中同时使用3×3和5×5两种卷积核,并将网络的前几层特征复用后进行加权,以提高重要特征的利用率。对10类牧草图像进行识别实验,结果表明,该方法识别率为94.1%,比VGG-13网络提高了5.7个百分点,双池化与多尺度特征加权有效提高了牧草识别精度。

Typical Forage Recognition Based on Double Pooling and Multi-scale Kernel Feature Weighted Convolution Neural Network

In order to solve the problem of forage recognition under natural conditions, a convolutional neural network method based on double-pooling feature weighting and multi-scale convolution kernel feature weighting structure was proposed. The spatial information and significance information of the image were fully utilized by using the dual-pooling feature weighted structure. Two groups of feature graphs were obtained by max-pooling and mean-pooling of feature graphs output from the convolution layer, and then these two groups of features were spliced. Finally, a feature re-calibration strategy was introduced to weight the importance of current tasks according to the feature graphs of each channel, so as to enhance useful features and suppress useless features. Image information was more fully mined by using multi-scale feature weighting structure. The 3×3 and 5×5 convolution kernels were used at the same time, and the features of the first several layers of the network were spliced with the features of the current layer to improve feature utilization rate. Feature re-calibration strategy was also introduced to weight features. The recognition experiments of ten pasture images showed that the recognition rate of the method was 94.1%, which was 5.7 percentage points higher than that of VGG-13 network, the double pooling and multi-scale feature weighting structure effectively improved the recognition accuracy.