基于深度全卷积神经网络的图像识别研究

Research on Image Recognition Based on Deep Fully Convolutional Neural Network

将要建立多层卷积网络模型,并使用AlexNet预训练模型,在此基础上进行迁移学习,使用kaggle的猫狗数据集进一步训练,模型最终能高灵活度、高准确率的识别猫狗图像,并且不受图像中猫狗的占比大小影响。该网络模型共有6 000万参数,一共包含8个卷积层,其中某些卷积层带有归一化层和池化层,最后一层是具有两个通道的图像输出,每个通道的值分别代表图像为猫和狗的概率。整个网络模型,弃用全连接层,选用全卷积网络来代替全连接层,大大提高网络的灵活性,解决了输入图像分辨率的限制问题,并且全卷积网络的前向传播更加高效,加快了训练的速度。为了方便分析以及进一步的研究,将可视化一层卷积和二层卷积所得到的卷积核和特征图。

A multi-layer convolutional network model will be established, and the AlexNet pre-training model will be used. On this basis, migration learning will be performed. Kaggle’s cat and dog data set will be used for further training. The model will eventually be able to recognize cat and dog images with high flexibility and high accuracy without being affected by the proportion of cats and dogs in the image. The network model has a total of 60 million parameters and a total of 8 convolutional layers. Some of the convolutional layers have a normalization layer and a pooling layer. The last layer is an image output with two channels, and the value of each channel respectively represent the probability whether the image is a cat or a dog. The fully connected layer is abandoned in the entire network model, and the fully connected layer is replaced by a fully convolutional network, which greatly improves the flexibility of the network, solves the limitation of input image resolution, and the forward propagation of the fully convolutional network is more efficient speeds up training. In order to facilitate analysis and further research, this article will visualize the convolution kernel and feature maps obtained by one-layer convolution and two-layer convolution.