Research On Lightweight Neural Network Data Compression Coding For Vision Terminal

Deep learning has achieved good results in many fields and has been widely used in various scenes of production and life.The success of deep learning is inseparable from the rapid development of computing chips in recent years,because deep neural networks have a huge amount of calculation.However,at the edge end,due to limited power consumption and computing ability,there are still huge challenges in deploying deep learning-based applications.In this research,aiming at the challenge of deep learning at the edge end,two lightweight neural network data compression coding methods are proposed.Recently,some researchers have tried to use information theory to open the black box of neural signal coding.In this research,inspired by the lossy data compression of wireless communication,the bitwise bottleneck method is proposed to quantize and encode neural network activation data.Based on the rate distribution theory,the bitwise bottleneck approach attempts to maintain the most important bits in activation representation by estimating the sparsity coefficient of different bits.Given a limited average code rate constraint,the bitwise bottleneck approach can minimize quantization distortion in a flexible layer-by-layer quantization manner.The experimental results on Image Net and other data sets show that by minimizing the quantization distortion of each layer,compared with the same period of domestic and foreign quantization research,the neural network with bitwise bottleneck layers achieved better classification accuracy under the low precision activation.Moreover,compared with the single-precision floating-point deep neural network,the bitwise bottleneck approach can improve the memory and calculation efficiency by 6.4 times and without affecting the accuracy of the neural network.On the other hand,in intelligent vision terminal equipments,not only the task of image recognition and classification is required,but also image compression is also needed.However,limited by the energy budget and computing resources of edge devices,it is still a big challenge to implement complex computing image compression algorithms.In response to this challenge,this research proposes a compressive convolutional network(CCN)method,which can learn isometric sensing matrices for compressive sensing from large data sets.In order to improve the incoherence between the sensing matrix and the sparse basis matrix,an incoherent regularization function is innovatively proposed.The proposed method reuses the convolution operator in the traditional object detection neural network without additional computational overhead.Experiments on several data sets show that the proposed image compression algorithm can achieve 3.1 times to 5.5 times higher compression efficiency than traditional approaches without affecting the original network detection accuracy,and the Image reconstruction PSNR is higher than the JPEG compression method and other compressive sensing methods with 2.7 dB to 5.2 dB.