Combining Learnable Low-dimensional Binary Filter Bases for Compressing Convolutional Neural Networks

Existing convolutional neural networks (CNNs) have achieved significant performance on various real-life tasks, but a large number of parameters in convolutional layers requires huge storage and computation resources which makes it difficult to deploy CNNs on memory-constraint embedded devices. In this paper, we propose a novel compression method that generates the convolution filters in each layer by combining a set of learnable low-dimensional binary filter bases. The proposed method designs more compact convolution filters by stacking the linear combinations of these filter bases. Because of binary filters, the compact filters can be represented using less number of bits so that the network can be highly compressed. Furthermore, we explore the sparsity of coefficient through L1-ball projection when conducting linear combination to avoid overfitting. In addition, we analyze the compression performance of the proposed method in detail. Evaluations on four benchmark datasets under VGG-16 and ResNet-18 structures show that the proposed method can achieve a higher compression ratio with comparable accuracy compared with the existing state-of-the-art filter decomposition and network quantization methods.

Combining Learnable Low-dimensional Binary Filter Bases for Compressing Convolutional Neural Networks

Existing convolutional neural networks (CNNs) have achieved significant performance on various real-life tasks, but a large number of parameters in convolutional layers requires huge storage and computation resources which makes it difficult to deploy CNNs on memory-constraint embedded devices. In this paper, we propose a novel compression method that generates the convolution filters in each layer by combining a set of learnable low-dimensional binary filter bases. The proposed method designs more compact convolution filters by stacking the linear combinations of these filter bases. Because of binary filters, the compact filters can be represented using less number of bits so that the network can be highly compressed. Furthermore, we explore the sparsity of coefficient through L1-ball projection when conducting linear combination to avoid overfitting. In addition, we analyze the compression performance of the proposed method in detail. Evaluations on four benchmark datasets under VGG-16 and ResNet-18 structures show that the proposed method can achieve a higher compression ratio with comparable accuracy compared with the existing state-of-the-art filter decomposition and network quantization methods.

The reversibility of one-dimensional cellular automata

Recently the reversible cellular automata are increasingly used to build high-performance cryptographic algorithms. The paper establishes a connection between the reversibility of homogeneous one-dimensional binary cellular automata of a finite size and the properties of a structure called binary filter with input memory and such finite automata properties as the prohibitions in automata output and loss of information. We show that finding the preimage for an arbitrary configuration of a one-dimensional cellular automaton of length L with a local transition function f is associated with reversibility of a binary filter with input memory. As a fact, the nonlinear filter with an input memory corresponding to our cellular automaton does not depend on the number of memory cells of the cellular automaton. The results obtained make it possible to reduce the complexity of solving massive enumeration problems related to the issues of reversibility of cellular automata. All the results obtained can be transferred to cellular automata with non-binary cell filling and to cellular automata of dimension greater than 1.

Cigarette Filter Rod Measure Based on Morphological Image Processing

Traditional touch measure method has the shortages of easily distorting cigarette filter rod, long measure time, not high measure precision and not online implementing. In order to overcome these, imaging measure technology is used to measure it. Firstly, preprocessing filtering is operated on the acquired image. Then sobel operator is used to detect the edge and binary filter rod target image is obtained through morphological processing after falsehood edge eliminating. After that correlation parameters are gained. Parameters measuring absolute error for two images acquired from the same filter rod under different illumination is 1.848%. Experiment proves that the proposed method can rapidly and correctly measure cigarette filter rod. It can effectively measure filter rod in practice.