基于卷积神经网络的智能冰箱果蔬图像识别的研究

Fruits and vegetables image recognition in intelligent refrigerator based on convolutional neural network

智能冰箱物体识别主要涉及对水果和蔬菜的识别,冰箱中果蔬数量繁多,光照不均,环境复杂,对此提出了一种用于处理该类识别问题的卷积神经网络。网络采用Re LU作为激活函数,它比传统的Sigmoid函数具有更强的稀疏能力和更大的梯度值,能够极大地加速网络收敛。隐含层中引入随机Dropout,使得某些节点不工作,减少节点间的"共同适应",降低网络对某一局部特征的过拟合,可减少网络计算复杂度并有效提升识别率。网络采用带动量项的基于梯度下降的反向传播算法,避免网络陷入局部极小值,提高识别率。最后通过用Supermarket Produce Dataset数据集模拟冰箱果蔬图像进行实验,验证了本文方法的有效性。

Object recognition in intelligent refrigerator mainly involves fruits and vegetables. Large stock of fruits and vegetables and uneven illumination make a complex environment in fridge. This paper presents a model of convolutional neural network(CNN) to solve such problems.The model uses Re LU as activation function which is stronger than Sigmoid with sparse ability and has larger gradient value,and it can greatly accelerate network convergence. Random Dropout is applied to the hidden layers to make some hidden units not work,which can reduce the phenomenon of ‘co-adaptation'between them. Besides,it can also reduce the possibility of overfitting to a local feature,which is able to simplify the algorithm complexity and improve recognition rate. The paper uses BP algorithm based on gradient descent with a momentum factor which could avoid the network falling into a local minimum value and enhance recognition rate. Finally,the supermarket produce dataset is used to simulate fruits and vegetables in fridge to identify the effectiveness of proposed methods.