Search Results (3)

On edge devices, convolutional neural network (CNN) inference is bottlenecked mainly by memory bandwidth, owing to the frequent memory accesses to feature maps and parameters. To address this challenge, we propose a memory-efficient hardware accelerator for depthwise separable convolution that minimizes off-chip memory traffic and parameter storage. The proposed architecture employs three key techniques: (1) a 64-bit run-length coding (RLC) packet compression that exploits feature-map sparsity after ReLU, (2) a mixed-precision scheme that represents feature maps and weights at different precisions, and (3) separated depthwise and pointwise convolution units. In particular, feature maps are transferred in RLC-compressed form, which reduces the amount of data exchanged with the host. The compressed data are decoded row by row, so the on-chip buffers hold only the rows required for computation rather than a complete feature map. In software simulation on ImageNet, the mixed-precision scheme reduced the parameter storage by 49.22% at a cost of 6.24 percentage points (pp) in Top-1 accuracy, and the RLC reduced the data by up to 56.07% in the deeper layers. Implemented on a Xilinx ZCU-104 FPGA, the proposed accelerator performs the depthwise separable convolution with a small number of logic resources and on-chip memory, confirming its feasibility for resource-constrained edge devices. Full article

Infrared line-scanning images have high redundancy and large file sizes. In JPEG2000 compression, the MQ arithmetic encoder’s complexity slows down processing. Huffman coding can achieve O(1) complexity based on a code table, but its integer-bit encoding mechanism and ignorance of the continuity of symbol distribution result in suboptimal compression performance. In particular, when encoding sparse quantized wavelet coefficients that contain a large number of consecutive zeros, the inaccuracy of the one-bit shortest code accumulates, reducing compression efficiency. To address this, this paper proposes Huf-RLC, a Huffman-based method enhanced with Run-Length Coding. By leveraging zero-run continuity, Huf-RLC optimizes the shortest code encoding, reducing the average code length to below one bit in sparse distributions. Additionally, this paper proposes a wavelet coefficient probability model to avoid the complexity of calculating statistics for constructing Huffman code tables for different wavelet subbands. Furthermore, Differential Pulse Code Modulation (DPCM) is introduced to address the remaining spatial redundancy in the low-frequency wavelet subband. The experimental results indicate that the proposed method outperforms JPEG in terms of PSNR and SSIM, while maintaining minimal performance loss compared to JPEG2000. Particularly at low bitrates, the proposed method shows only a small gap with JPEG2000, while JPEG suffers from significant blocking artifacts. Additionally, the proposed method achieves compression speeds 3.155 times faster than JPEG2000 and 2.049 times faster than JPEG. Full article

Congestion in a wireless sensor network causes an increase in the amount of data loss and delays in data transmission. In this paper, we propose a new congestion control technique (ACT, Adaptive Compression-based congestion control Technique) based on an adaptive compression scheme for packet reduction in case of congestion. The compression techniques used in the ACT are Discrete Wavelet Transform (DWT), Adaptive Differential Pulse Code Modulation (ADPCM), and Run-Length Coding (RLC). The ACT first transforms the data from the time domain to the frequency domain, reduces the range of data by using ADPCM, and then reduces the number of packets with the help of RLC before transferring the data to the source node. It introduces the DWT for priority-based congestion control because the DWT classifies the data into four groups with different frequencies. The ACT assigns priorities to these data groups in an inverse proportion to the respective frequencies of the data groups and defines the quantization step size of ADPCM in an inverse proportion to the priorities. RLC generates a smaller number of packets for a data group with a low priority. In the relaying node, the ACT reduces the amount of packets by increasing the quantization step size of ADPCM in case of congestion. Moreover, in order to facilitate the back pressure, the queue is controlled adaptively according to the congestion state. We experimentally demonstrate that the ACT increases the network efficiency and guarantees fairness to sensor nodes, as compared with the existing methods. Moreover, it exhibits a very high ratio of the available data in the sink. Full article