Optimal Selection of Pyramid Pooling Components for Convolutional Neural Network Classifier
Siti Raihanah Abdani, Mohd Asyraf Zulkifley · 2020 International Conference on Decision Aid Sciences and Application (DASA) · 2020
There has been an increasing number of researches on deep learning applications in screening eye diseases. Since most of the screening tools will be deployed for rural areas that lack modern medical equipment, it is best if the screening algorithm can be deployed on a mobile platform. Thus, a lightweight deep learning model very much fits the mobile platform as it requires low memory storage and imposes a low computational burden on the hardware. Therefore, a modified ShuffleNet V1 network is proposed to screen eye diseases using fundus images. A spatial pyramid pooling module is integrated at the exit flow of the network, such that the memory usage remains relatively the same but with improved classification accuracy. The better performance can be attributed to the exit module that better extracts the features from various scales rather than a simple global average pooling operator. The best mean accuracy of 0.7564 is obtained when maximum down-pooling operators are used with a kernel set of 2, 4, and 7. Furthermore, a wrong selection of hyper-parameters such as the case of kernel set 5, 6, and 7 also with maximum pooling operators return the lowest accuracy of just 0.7011. Therefore, an optimal selection of kernel set and down pooling operators is vital in improving classification performance. The proposed lightweight model can be further improved by using a separable convolution scheme, which a factorized version of a regular convolution.