基于多尺度特征融合的水果图像识别算法研究

Research on Fruit Image Recognition AlgorithmBased on Multiscale Feature Fusion

水果图像识别是智能采摘系统中最重要的组成部分。针对现阶段水果图像识别过程中存在的漏检和误检现象,为进一步提高识别准确率,研究了基于多尺度特征融合的水果图像识别算法。首先,为避免训练过程中出现欠拟合现象,对Fruits-360中的水果图像进行数据扩充,并进一步灰度归一化处理以减少计算量。随后采用ResNet-50作为骨干网络,并在骨干网络的基础上建立多尺度采样层,使用1×1、3×3和5×5的卷积核在拓宽网络宽度的同时进行特征提取,多尺度网络层整体增加BN层,即在每个卷积层之后都增加BN层。使在ResNet-50提取的原始特征基础上获取语义信息更加丰富的特征图。最后采用梯度下降法优化网络,得到最终的识别模型。实验结果表明,所提算法识别精度高,可准确的对水果图像经识别,识别精度高达99.4%,在相同数据集的情况下,优于目前主流算法,可为水果自动采摘技术提供帮助。

Fruit image recognition is the most important component of intelligent picking systems.To further improve the recognition accuracy,the fruit image recognition algorithm based on a multi-scale feature fusion was studied given the phenomenon of missed detection and false detection in the current fruit image recognition process.First,to avoid the under-fitting phenomenon during the training process,the fruit image in Fruits-360 is expanded,and further grayscale normalization is performed to reduce the amount of calculation.Then,Resnet-50 is used as the backbone network,and a multi-scale sampling layer is built on the backbone network.The convolution kernels of 1×1,3×3 and 5×5 are used to extract the features while widening the network width.The multi-scale network layer increases the BN layer as a whole.That is,the BN layer is added after each convolutional layer.A feature map rich in semantic information is obtained based on the original features extracted by Resnet-50.Finally,the gradient descent method is used to optimize the network to obtain the final recognition model.The experimental results show that the high precision of the proposed algorithm can accurately identify the fruit image,and the recognition accuracy is up to 99.4%.In the case of the same data set,it is better than the current mainstream algorithm and can help the automatic fruit picking technology.