Cascaded pooling for Convolutional Neural Networks

Despite of its increasing popularity, Deep Convolutional Neural Network (DCNN) fails to achieve accurate image classification results due to overfitting problem. Overfitting occurs due to lack of enough training data which are used to optimize a large number of parameters of DCNN. Dropout helps to solve this problem but it works only on fully connected layers. As pooling layer helps in reducing high dimensional convolutional features, our main focus is to develop an efficient pooling method which retains the most distinguishable features. In this work, we proposed a pooling approach called cascaded pooling which consists of two pooling methods: max-max pooling and max-average pooling. We evaluated our approach on InceptionV3 architecture of DCNN where the first two max pooling layers are replaced by max-max pooling and the last average pooling layer by max-average pooling and obtained improved classification results on two aerial scene datasets.