基于改进深度卷积神经网络的苹果病害识别

Apple disease identification based on improved deep convolutional neural network

传统的深度卷积神经网络(DCNNs)使用3个全连接层将经过多次卷积层和池化层后提取到的特征图映射并连接为一个特征向量,然后利用Softmax分类器进行分类。该模型容易出现过拟合问题,而且由于在全连接层中参数太多,导致训练时间增加和泛化能力下降。针对传统的DCNNs模型在图像识别中出现的问题,提出一种改进的DCNNs模型,并应用于苹果叶部病害识别中。相比传统的DCNNs算法,改进的DCNNs利用一个全局平均池化层替代全连接层,并利用改进的Softmax分类器进行病害类别识别。在苹果病害叶片图像数据库上的实验结果表明,该模型能够克服过拟合问题,提高病害的识别率,大幅度降低模型的训练和识别时间。

Deep convolutional neural networks(DCNNs)have achieved impressive performance in the field of image classification.In the classical DCNNs,there are often two or three fully connected layers to be used to map the learned feature maps and concatenate them as a one-dimensional feature vector,and then are input into the Softmax classifier to classify the images.The drawbacks of these DCNNs models are the over-fitting phenomenon,long time to train their parameters and weak generalization ability due to too many parameters in the fully connected layers,especially in the last fully connected layer.Because there are too many parameters in the model and limited training samples in practice,the trained model is easy to produce over-fitting.Over-fitting problem can be solved by Dropout.Because the weight update obtained by Dropout is no longer dependent on the joint action of hidden nodes with fixed relations,preventing certain features from having effects only under other specific features,therefore the network can learn more robust features.Although dropout is effective to handle the over-fitting,one of the disadvantages is that the cost function is no longer clearly defined.Because in iterations,some points are randomly removed,it is hard to check the gradient descent performance.In order to improve the performance of DCNNs-based image recognition method,an improved DCNNs(IDCNNs)model is proposed and applied to apple disease identification.Different from the classical DCNNs,the proposed model employs one global average pooling layer instead of all the full connection layers and utilizes the modified Softmax classifier replacing the Softmax classifier of DCNNs.IDCNNs have two advantages.One is that there are few parameters needing to be optimized in the global pooling layer,thus over-fitting is naturally avoided at this layer.The other is that the global pooling sums up the spatial information,hence it is more robust to spatial translation of the input images.Moreover,IDCNNs is more consistent with the working structure of DCNNs.Each feature map is correlated with the category output,rather than the unit of the feature map is directly correlated with the category output.Experimental results on an actual apple diseased leaf image database show that the proposed method can address the overfitting problem,improve the recognition rate of disease recognition and greatly reduce the training and recognition time of the model.