Enhancing Convolutional Neural Network Performance Using Domain Transforms in Constrained Networks

Convolutional neural network (CNN) has emerged as one of the primary tools for image classification. In particular, deep CNNs are now considered the main tool for this purpose. However, these networks are often large and require computing and storage power that may not be available in very small sensor devices such as IoT (Internet of Things) devices; their training is also time and computing power consuming. As a result, in some applications, reducing the size of input data (images) and the processing network becomes necessary. Such reduction usually comes at the cost of reduced classification performance. In this paper, we consider networks with under 200k learnable parameters, as opposed to millions in deeper networks. We examine how domain transforms can be used for efficient size reduction and improvement of classification accuracy for small networks. We emphasize that finding optimal hyperparameters or network configurations is not our objective in this paper. It is shown that by using transforms such as discrete wavelet transforms (DWT) or discrete cosine transform (DCT), it is possible to efficiently improve the performance of size-reduced networks and inputs. We demonstrate that in most cases, the improvement can be traced to higher entropy of resized input using transforms. While transforms such as DCT allow variable input and network sizes to be utilized, DWT proves to be very effective when significant size reduction is needed (improving the result by up to 5%). It is also shown that input size reduction of up to 75% is possible, without loss of classification accuracy in some cases. We use two datasets of small images, including Fashion MNIST and CIFAR-10, to evaluate the performance of size reduction methods.