Concatenated Global Average Pooled Deep Convolutional Embedded Clustering

Abstract

Deep Clustering learns cluster friendly salient features in embedded space. In our previous work of Global Average Pooled Deep Convolutional Embedded Clustering (GAPDCEC) algorithm, the last convolution layer feature maps are pooled to build the embedded space. This considers only spatial information retains in the last convolution layer of the encoder, which unable to capture discriminative features from entire convolutional layers. To address this issue, we propose a solution using concatenation of all convolutional layer outputs and then Global Average Pooling (GAP) is applied on the concatenated feature maps in the encoder. This will encourage the network to learn cluster friendly features of all convolutional layers. Our experimental results prove the efficiency of proposed Concatenated Global Average Pooled Deep Convolutional Embedded Clustering (CGAPDCEC).