首页期刊导航|Journal of signal processing systems for signal, image, and video technology
期刊信息/Journal information
Journal of signal processing systems for signal, image, and video technology
Springer Science + Business Media
Springer Science + Business Media
月刊
1939-8018
Journal of signal processing systems for signal, image, and video technology/Journal Journal of signal processing systems for signal, image, and video technologySCIISTP
查看更多>>摘要：Recent smartphones employ multi-camera setups for capturing images, prompting the exploration of stereo image super-resolution (SSR) algorithms. SSR uses the complementary information provided by a binocular system to upscale input stereo image pairs. The effectiveness of SSR algorithms depends on successfully utilizing the stereo information from the training images. This paper, proposes a lightweight stereo image super-resolution method using modified parallax attention (LmPASSR), which enhances the utilization of stereo information. This is achieved through a modified occlusion mask that filters out irrelevant attention values. Additionally, the model incorporates depth-wise convolutions, implemented as D-blocks, to minimize parameter usage. Experimental results demonstrate that despite having fewer parameters, the proposed model produces results comparable to state-of-the-art (SOTA) methods.
查看更多>>摘要：Polar codes are a class of linear error-correction codes that have received a lot of attention due to their ability to achieve channel capacity in an arbitrary binary discrete memoryless channel (B-DMC) with low-complexity successive-cancellation (SC) decoding. However, practical implementations often require better error-correction performance than what SC decoding provides, particularly at short to moderate code lengths. Successive-cancellation flip (SCF) decoding algorithm was proposed to improve error-correction performance with an aim to detect and correct the first wrongly estimated bit in a codeword before resuming SC decoding. At each additional SC decoding trial, i.e., decoding attempt beyond the initial unsuccessful trial, one bit estimated as the least reliable is flipped. Dynamic SCF (DSCF) is a variation ofSCF, where multiple bits may be flipped simultaneously per trial. Despite the improved error-correction performance compared to the SC decoder, SCF-based decoders have variable execution time, which leads to high average execution time and latency. In this work, we propose the generalized restart mechanism (GRM) that allows to skip decoding computations that are identical between the initial trial and any additional trial. Under DSCF decoding with up to 3-bit flips per decoding trial, our proposed GRM is shown to reduce the average execution time by 25% to 60% without any negative effect on error-correction performance. The proposed mechanism is adaptable to state-of-the-art latency-reduction techniques. When applied to Fast-DSCF-3 decoding, the additional reduction brought by the GRM is 15% to 22%. For the DSCF-3 decoder, the proposed mechanism requires approximately 4% additional memory.
查看更多>>摘要：Vision Transformer (ViT) has shown great promise in image processing. However, its large model parameters and computation complexity result in inference delays, making deployment on edge devices challenging. To overcome these issues, various model compression techniques like quantization and distillation have been developed. Previous studies have explored quantization and binarization of ViT, but their effectiveness in minimizing accuracy loss has been limited, and primarily focusing on software solutions. Research on hardware acceleration remains underexplored but is essential for boosting the inference speed of binarized networks. This paper proposes a hardware acceleration scheme for binarized ViT using a distribution matching layer. Our approach starts with an experimental and theoretical analysis of the data distribution in binarized ViT, leading to the introduction of a distribution matching layer post-binarization. We also design a compatible model storage scheme and a hardware acceleration algorithm to enhance the efficiency of weight matrix storage and computation. Additionally, optimizing large matrix multiplication within the self-attention layer significantly improves overall model speed. Experimental results show that our method increases accuracy by 10% compared to traditional binarized ViT approaches with learning factors, reducing the accuracy gap between binarized and full-precision models to 4%. Furthermore, our approach achieves inference speeds approximately 45 times faster than traditional models.
查看更多>>摘要：With the rapid development of blockchain technology, the role of smart contracts as an important part of the blockchain has become more and more significant, bringing unprecedented value and innovation to many fields. However, despite the immense value created by smart contracts, their potential vulnerabilities have led to numerous attacks, resulting in substantial financial losses. Conventional expert-based detection methods, along with machine learning and deep learning techniques, frequently face challenges such as low accuracy and insufficient reliability. In order to address these issues, this paper puts forward a novel deep learning vulnerability detection method based on opcode-level analysis, designated as NDLSC. The method initially transforms smart contracts into opcodes, subsequently employing the Skip-Gram model in Word2Vec to vectorise the dataset. Subsequently, the Residual Networks 34(ResNet-34) deep learning model is utilized for feature extraction, followed by the Kolmogorov-Arnold Networks(KAN) model for further feature extraction and classification. This approach is employed with the objective of achieving superior results. The core algorithm of NDLSC, which combines ResNet and KAN, is experimentally compared with existing vulnerability detection techniques. The findings demonstrate that this combination not only enhances the precision of smart contract vulnerability identification but also fortifies the resilience of the model. By organically combining these two structures, the understanding and detection of smart contracts are significantly improved, making the detection process more precise and reliable.
查看更多>>摘要：The real-time engineering applications concerning audio and image compressions necessitate high-performance VLSI architectures of discrete wavelet transform (DWT). This paper suggests efficient folded VLSI architectures of convolution based 1D/2D-DWT utilizing bit accumulation. In the proposed folded convolution based fixed point 1 D/2D-DWTs, quarter precision Wallace tree multiplier and multiply accumulate circuit (MAC) are incorporated, which accumulate the resultant values in every level of decomposition of the DWT. Here, the n × n-bit quarter precision multiplier and MAC are utilized to execute an n x n-bit or two numbers of n × n/2-bit or four numbers of n/2 × n/2-bit multiplications and MAC operations concurrently, respectively. To enhance the value of peak signal to noise ratio (PSNR) in audio and image compressions, the rounding off operation is performed only at the final level of DWT instead of at each level. Cadence is used to implement all current and proposed DWT designs with 45 nm CMOS technology.
查看更多>>摘要：Denoising enhances image quality by separating noise from observed signals, eliminating extraneous information while preserving essential features and image integrity. However, existing surveys on conventional denoising techniques often focus solely on processing-domain taxonomies, thereby neglecting evolutionary relationships, overlooking recent advances, and lacking multi-modal exploration. Consequently, modern machine learning pipelines have not fully exploited classical techniques. To advance multi-modal denoising and inspire new learning-based algorithms, this paper presents a comprehensive review of traditional denoising methods, quantitatively assesses their cross-modal transferability, and explores their integration into learning frameworks. Specifically, (1) this work proposes a novel taxonomy for conventional denoising techniques, including domain-based and signal-decomposition-based approaches, provides a systematic analysis of their evolutionary relationships, and investigates recent advances. (2) The study evaluates multi-modal denoising performance by applying baseline methods to infrared images and conducting a comparative analysis. (3) This paper surveys the latest research on traditional approaches, retraces their co-evolution with machine learning, and specifically explores the potential for fusing these techniques within learning-based denoising algorithms. In general, this review serves as a valuable reference for researchers in RGB-infrared denoising, image restoration, and related fields. The advancements in these areas significantly impact various domains, including defect detection in industrial production, worker protection safety recognition, and object tracking in smart transportation.
NETL
NSTL
Springer Nature