Author Email: firstname.lastname@example.org
In the existing convolutional neural networks, the majority of the used pooling operations are max pooling or mean pooling, but it would lose some important feature information when processing the feature maps. Here we report interpolation pooling to overcome the problem for retaining more effective information of feature maps. The interpolation pooling takes the known pixel points of 4×4 with the nearest to the interpolation point into account. Due to the distance from the pixels to be inserted, the weight of the pixels near the distance in the calculation is larger. We apply it to different convolutional neural networks, such as lenet-5 and pyramid convolutional neural networks. We found that the method has the advantages of faster convergence and higher accuracy than the traditional method of pooling.
Interpolation Pooling, Image Classification, Convolutional Networks
In recent years, deep learning has attracted the attention of many scholars. The fact has proved that deep learning has deeper and wider application than traditional shallow learning networks, including visual recognition, speech recognition and natural language processing. In 2006, Hinton improved the method(deep belief nets) of deep learning breaks the bottleneck of the development of BP neural network. In all kinds of deep learning, convolutional neural networks (CNNs) have been the most extensively studied. CNNs consist of three types of layers: convolution, pooling and fully connected layer. For convolutional layers, the convolution kernel is shared by all the spatial positions. which reduce the complexity of the model and make the network easier to train. Pooling is an important concept of CNNs, including max pooling, mean pooling or mixed pooling. A pooling layer reduces computational load by reducing the number of convolutional layers. In 2012, Krizhevsky et al. proposed an AlexNet model that shows significant improvements. AlexNet is similar to LeNet-5, but with a deeper structure. Simonyan et proposed the VGG network based on AlexNet. And he proved that the enhancement of net work depth helps to improve the accuracy of image classification. By increasing the depth, the network can better approximate the objective function, increase the non-linearity, and get a better representation of the features. However, this also increases the complexity of the network and makes it more difficult to optimize. To solve degradation problem with increasing the depth of CNNs, He et proposed a ResNet that won the 2015 ILSVRC championship. ResNet maps lowlevel features directly to high-level Network. And it is eight times as deep as VGG and 20 times faster than AlexNet. Szegedy et. proposed an inception module by observing and optimizing the network structure, which reduces the network complexity and replaces the previous convolution kernel by using a 1 × 1 convolution kernel in the inception module. The number of training parameters for GoogLeNet built using the Inception module is only 1 / 12th of AlexNet, but the accuracy of image classification on ImageNet is improved. In 2017, Saining Xie proposed the ResNeXt network structure based on ResNet. ResNeXt improves the accuracy without increasing the complexity of the parameters and reducing the number of hyper-parameters. At the same time, a variety of methods [9-11]have been proposed to overcome the difficulties encountered in deep CNNs training. All the methods mentioned above are improvements for the depth, activation function and convolution kernel of CNNs. In these models, max pooling or mean pooling is used. Max pooling simply selects the maximum value from the pooling area as the final response value, which is sensitive to noise information. Mean pooling takes the average values in the pooling area, which effectively reduces the impact of noise information, but it smooths the image and leads to the loss of high frequency information.In this paper, for the existing problems of the max pooling and the mean pooling in CNNs, interpolation pooling is proposed to optimize the network efficiency. Interpolation pooling mainly uses the method of image interpolation to select the nearest 16 pixels as the pixel value corresponding to the final image, so as to achieve the purpose of scaling the feature map.
In the framework of convolutional neural networks, this paper proposes interpolation pooling strategies that can be used in combination with other methods. The application of interpolation pooling can effectively prevent the occurrence of over-fitting and effectively improve the robustness of the model while retaining the high frequency information and effectively preventing the loss of information such as edges. In addition, interpolation pooling can be combined with any existing convolution neural network model. Experiments show that this method can effectively improve the convergence rate of the model while ensuring the accuracy of the model.
559 total views, 1 views today