A pilot study of novel multi-filter CNN layer.

Abstract

Convolutional neural networks (CNNs) have reached their peak of complex structures, but until now, few researchers have addressed the problem of relying on one filter size. Mainly a 3 × 3 filter is the most common one used in any structure. Only at the first layers of the CNN model, filters bigger than 3 × 3 could be partially used. Most researchers work with filters (size, values, etc.) as a black box. To the best of our knowledge, this research is the first pilot study that proposes a new multi-filter layer in which different filters with variant sizes are used to replace the 3 × 3 filter layers. Our proposed multi-filter layer has yielded encouraging results, demonstrating notable improvements ranging from 1% to 5% in performance. This achievement was realized by developing two innovative structures, namely the fixed structure and the decreasing structure. Both of them leverage the multi-filter layer. Although the two structures exhibit promising outcomes, the later structure offers the additional advantages of reduced computational requirements and enhanced learner strength.