Tamer Shanableh
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Tamer Shanableh has a total of 41 co-authors. Here are the top co-authors by number of shared publications:
- Khaled Assaleh
American University of Sharjah
3 shared publications - Obada Issa
Department of Computer Science and Engineering, American University of Sharjah, Sharjah P.O. Box 26666, United Arab Emirates
3 shared publications K. Assaleh
Department of Electrical Engineering, American University of Sharjah
2 shared publicationsMahitab Hassan
IBM Cloud, Dubai, UAE
2 shared publications- Yaman Albadawi
American University of Sharjah,Department of Computer Science and Engineering,Sharjah,United Arab Emirates
2 shared publications - A. R. Al-Ali
Department of Computer Science and Engineering, American University of Sharjah, Sharjah, United Arab Emirates
1 shared publication
Latest Publications
- 14 August 2026
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 85 (8) , 700
Temporal domain encoding with TCN-BiLSTM and late fusion for continuous Arabic sign language recognition in signer-independent mode
- 14 August 2026
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 85 (8) , 700
Temporal domain encoding with TCN-BiLSTM and late fusion for continuous Arabic sign language recognition in signer-independent mode
- Open Access
- 3 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 85635-85651
Lite-SignVLM: An Efficiency-Oriented Framework for Sign Language Recognition With Temporal Storyboards
Sign language recognition (SLR) systems have made substantial progress in recent years; however, practical deployment remains challenging in real-world scenarios that require low latency, offline operation, and execution on resource-constrained devices. Although large transformer-based and vision-language models achieve strong recognition accuracy, their reliance on computationally expensive backbones leads to high memory usage and inference cost. Additionally, critical efficiency metrics such as model size, GFLOPs, and inference latency are often unreported. In this work, we address these limitations by proposing a comprehensive, efficiency-oriented compression framework for SLR. Starting from a high-accuracy teacher model such as SignVLM, we distill knowledge into lightweight student models using three complementary strategies. First, we introduce a storyboard-based video representation that encodes the full temporal progression of a sign into a single structured RGB image. Second, we perform systematic architectural distillation by comparing CLIP and EfficientNet backbones under different decoder retention rates. Third, we conduct a detailed study of post-training FP16 casting and INT8 quantization. Experiments on the KArSL-502 dataset under a signer-dependent evaluation protocol show that preserving full temporal input motion enables student models to match or exceed teacher accuracy despite significantly reduced complexity. Our best configuration achieves 99.34% Top-1 accuracy with only 5.38M parameters, 15.45 MB disk size, and 0.43 GFLOPs. While FP16 casting consistently preserves accuracy across architectures, INT8 quantization exhibits architecture-dependent behavior in which models with EfficientNet-based backbones maintain high recognition accuracy, whereas models with CLIP-based backbones suffer from severe recognition degradation due to outlier activation effects. Overall, this work provides a clear and deployable pathway for on-device SLR systems, bridging the gap between high recognition performance and practical efficiency constraints.
- 3 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 85635-85651
Open AccessLite-SignVLM: An Efficiency-Oriented Framework for Sign Language Recognition With Temporal Storyboards
Sign language recognition (SLR) systems have made substantial progress in recent years; however, practical deployment remains challenging in real-world scenarios that require low latency, offline operation, and execution on resource-constrained devices. Although large transformer-based and vision-language models achieve strong recognition accuracy, their reliance on computationally expensive backbones leads to high memory usage and inference cost. Additionally, critical efficiency metrics such as model size, GFLOPs, and inference latency are often unreported. In this work, we address these limitations by proposing a comprehensive, efficiency-oriented compression framework for SLR. Starting from a high-accuracy teacher model such as SignVLM, we distill knowledge into lightweight student models using three complementary strategies. First, we introduce a storyboard-based video representation that encodes the full temporal progression of a sign into a single structured RGB image. Second, we perform systematic architectural distillation by comparing CLIP and EfficientNet backbones under different decoder retention rates. Third, we conduct a detailed study of post-training FP16 casting and INT8 quantization. Experiments on the KArSL-502 dataset under a signer-dependent evaluation protocol show that preserving full temporal input motion enables student models to match or exceed teacher accuracy despite significantly reduced complexity. Our best configuration achieves 99.34% Top-1 accuracy with only 5.38M parameters, 15.45 MB disk size, and 0.43 GFLOPs. While FP16 casting consistently preserves accuracy across architectures, INT8 quantization exhibits architecture-dependent behavior in which models with EfficientNet-based backbones maintain high recognition accuracy, whereas models with CLIP-based backbones suffer from severe recognition degradation due to outlier activation effects. Overall, this work provides a clear and deployable pathway for on-device SLR systems, bridging the gap between high recognition performance and practical efficiency constraints.
- Open Access
- 13 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 57510-57520
Ghost-NeRV: Efficient Neural Video Representation via Ghost Convolutions
Neural video representation (NeRV) has emerged as an efficient paradigm for video compression by encoding entire sequences into neural network parameters. Despite its strong reconstruction capability, NeRV suffers from high computational cost due to expensive convolutional operations in the decoder, limiting its applicability in resource-constrained environments. In this work, we propose Ghost-NeRV, an efficient extension of NeRV that reduces computational complexity through architectural optimization rather than numerical approximation. The work presents a systematic architectural efficiency study of structural redundancy within NeRV decoders, quantifying the trade-off between computational complexity, model compactness, and reconstruction fidelity. The proposed Ghost-NeRV integrates Ghost convolutions into the NeRV decoder to exploit feature redundancy and generate expressive representations using low-cost operations. Unlike binarization-based approaches, Ghost-NeRV preserves full-precision computation while significantly reducing the number of floating-point operations and model parameters. Extensive experiments on multiple video sequences demonstrate that Ghost-NeRV achieves up to 50% reduction in GFLOPs and approximately 19% reduction in model size, while maintaining stable reconstruction quality. Compared to the NeRV baseline, Ghost-NeRV incurs only a modest degradation in PSNR (typically within 1 dB) and preserves temporal consistency, significantly outperforming Binary-NeRV in perceptual stability. An ablation study further evaluates a GhostConv-V2 variant, which provides marginal quality improvements at the cost of increased computation and bitrate, confirming that the original Ghost convolution offers the best efficiency-accuracy trade-off. These results demonstrate that Ghost-NeRV provides an effective and practical solution for neural video representation, enabling substantial computational savings while maintaining high perceptual quality, improving its suitability for resource-constrained environments.
- 13 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 57510-57520
Open AccessGhost-NeRV: Efficient Neural Video Representation via Ghost Convolutions
Neural video representation (NeRV) has emerged as an efficient paradigm for video compression by encoding entire sequences into neural network parameters. Despite its strong reconstruction capability, NeRV suffers from high computational cost due to expensive convolutional operations in the decoder, limiting its applicability in resource-constrained environments. In this work, we propose Ghost-NeRV, an efficient extension of NeRV that reduces computational complexity through architectural optimization rather than numerical approximation. The work presents a systematic architectural efficiency study of structural redundancy within NeRV decoders, quantifying the trade-off between computational complexity, model compactness, and reconstruction fidelity. The proposed Ghost-NeRV integrates Ghost convolutions into the NeRV decoder to exploit feature redundancy and generate expressive representations using low-cost operations. Unlike binarization-based approaches, Ghost-NeRV preserves full-precision computation while significantly reducing the number of floating-point operations and model parameters. Extensive experiments on multiple video sequences demonstrate that Ghost-NeRV achieves up to 50% reduction in GFLOPs and approximately 19% reduction in model size, while maintaining stable reconstruction quality. Compared to the NeRV baseline, Ghost-NeRV incurs only a modest degradation in PSNR (typically within 1 dB) and preserves temporal consistency, significantly outperforming Binary-NeRV in perceptual stability. An ablation study further evaluates a GhostConv-V2 variant, which provides marginal quality improvements at the cost of increased computation and bitrate, confirming that the original Ghost convolution offers the best efficiency-accuracy trade-off. These results demonstrate that Ghost-NeRV provides an effective and practical solution for neural video representation, enabling substantial computational savings while maintaining high perceptual quality, improving its suitability for resource-constrained environments.
- Open Access
- 20 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Ghost-NeRV: Efficient Neural Video Representation via Ghost Convolutions
Neural video representation (NeRV) has emerged as an efficient paradigm for video compression by encoding entire sequences into neural network parameters. Despite its strong reconstruction capability, NeRV suffers from high computational cost due to expensive convolutional operations in the decoder, limiting its applicability in resource-constrained environments. In this work, we propose Ghost-NeRV, an efficient extension of NeRV that reduces computational complexity through architectural optimization rather than numerical approximation. Ghost-NeRV integrates Ghost convolutions into the NeRV decoder to exploit feature redundancy and generate expressive representations using low-cost operations. Unlike binarization-based approaches, Ghost-NeRV preserves full-precision computation while significantly reducing the number of floating-point operations and model parameters. Extensive experiments on multiple video sequences demonstrate that Ghost-NeRV achieves up to 50% reduction in GFLOPs and approximately 19% reduction in model size, while maintaining stable reconstruction quality. Compared to the NeRV baseline, Ghost-NeRV incurs only a modest degradation in PSNR (typically within 1 dB) and preserves temporal consistency, significantly outperforming Binary-NeRV in perceptual stability. An ablation study further evaluates a GhostConv-V2 variant, which provides marginal quality improvements at the cost of increased computation and bitrate, confirming that the original Ghost convolution offers the best efficiency-accuracy trade-off. These results demonstrate that Ghost-NeRV provides an effective and practical solution for neural video representation, enabling substantial computational savings while maintaining high perceptual quality, making it well-suited for deployment in resource-constrained video compression scenarios.
- 20 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Open AccessGhost-NeRV: Efficient Neural Video Representation via Ghost Convolutions
Neural video representation (NeRV) has emerged as an efficient paradigm for video compression by encoding entire sequences into neural network parameters. Despite its strong reconstruction capability, NeRV suffers from high computational cost due to expensive convolutional operations in the decoder, limiting its applicability in resource-constrained environments. In this work, we propose Ghost-NeRV, an efficient extension of NeRV that reduces computational complexity through architectural optimization rather than numerical approximation. Ghost-NeRV integrates Ghost convolutions into the NeRV decoder to exploit feature redundancy and generate expressive representations using low-cost operations. Unlike binarization-based approaches, Ghost-NeRV preserves full-precision computation while significantly reducing the number of floating-point operations and model parameters. Extensive experiments on multiple video sequences demonstrate that Ghost-NeRV achieves up to 50% reduction in GFLOPs and approximately 19% reduction in model size, while maintaining stable reconstruction quality. Compared to the NeRV baseline, Ghost-NeRV incurs only a modest degradation in PSNR (typically within 1 dB) and preserves temporal consistency, significantly outperforming Binary-NeRV in perceptual stability. An ablation study further evaluates a GhostConv-V2 variant, which provides marginal quality improvements at the cost of increased computation and bitrate, confirming that the original Ghost convolution offers the best efficiency-accuracy trade-off. These results demonstrate that Ghost-NeRV provides an effective and practical solution for neural video representation, enabling substantial computational savings while maintaining high perceptual quality, making it well-suited for deployment in resource-constrained video compression scenarios.
- Open Access
- 28 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 16143-16157
Binary-NeRV: Hybrid-Precision Weights Binarization for Efficient Neural Video Representation
Neural implicit video representations such as NeRV have emerged as a powerful alternative to traditional video codecs. However, the high computational cost and full-precision storage of NeRV limit its practicality for resource-constrained and embedded platforms. In this work, we propose Binary-NeRV, a hybrid-precision extension of NeRV that integrates XNOR-based binary convolutions into the decoding pipeline to significantly reduce model size, bitrate, and computational complexity while reasonably preserving reconstruction fidelity. Inspired by XNOR-Net, convolutional weights are binarized to ±1 with learned scaling factors, while critical components such as the stem MLP, normalization layers, activations, skip connections, and upsampling operators are selectively retained in FP32 to ensure stable training and high visual quality. We introduce two directional progressive binarization strategies, Left-to-Right (L2R) and Right-to-Left (R2L), to analyze the impact of binarizing layers at different spatial resolutions. A detailed complexity analysis shows that over 75% of NeRV’s computational cost is dominated by the final high-resolution convolutional layer, enabling highly effective targeted binarization. Extensive experiments on standard video sequences demonstrate that selectively binarizing only the deepest layer achieves up to 68–89% reduction in equivalent GFLOPs with reasonable degradation in PSNR and MS-SSIM. Temporal consistency analysis shows that selective binarization preserves temporal stability, while aggressive full-component binarization introduces noticeable flickering artifacts. We additionally present an ablation study where all convolutional layers are binarized, both before and after pruning and quantization. These results systematically validate the hybrid-precision design, showing that full binarization yields substantial bitrate reductions but at a notable quality cost, thereby justifying the selective binarization strategy adopted in Binary-NeRV. Compared with state-of-the-art NeRV variants, Binary-NeRV delivers substantial efficiency gains, establishing hybrid binarization as a practical and scalable approach for efficient neural video representation.
- 28 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 16143-16157
Open AccessBinary-NeRV: Hybrid-Precision Weights Binarization for Efficient Neural Video Representation
Neural implicit video representations such as NeRV have emerged as a powerful alternative to traditional video codecs. However, the high computational cost and full-precision storage of NeRV limit its practicality for resource-constrained and embedded platforms. In this work, we propose Binary-NeRV, a hybrid-precision extension of NeRV that integrates XNOR-based binary convolutions into the decoding pipeline to significantly reduce model size, bitrate, and computational complexity while reasonably preserving reconstruction fidelity. Inspired by XNOR-Net, convolutional weights are binarized to ±1 with learned scaling factors, while critical components such as the stem MLP, normalization layers, activations, skip connections, and upsampling operators are selectively retained in FP32 to ensure stable training and high visual quality. We introduce two directional progressive binarization strategies, Left-to-Right (L2R) and Right-to-Left (R2L), to analyze the impact of binarizing layers at different spatial resolutions. A detailed complexity analysis shows that over 75% of NeRV’s computational cost is dominated by the final high-resolution convolutional layer, enabling highly effective targeted binarization. Extensive experiments on standard video sequences demonstrate that selectively binarizing only the deepest layer achieves up to 68–89% reduction in equivalent GFLOPs with reasonable degradation in PSNR and MS-SSIM. Temporal consistency analysis shows that selective binarization preserves temporal stability, while aggressive full-component binarization introduces noticeable flickering artifacts. We additionally present an ablation study where all convolutional layers are binarized, both before and after pruning and quantization. These results systematically validate the hybrid-precision design, showing that full binarization yields substantial bitrate reductions but at a notable quality cost, thereby justifying the selective binarization strategy adopted in Binary-NeRV. Compared with state-of-the-art NeRV variants, Binary-NeRV delivers substantial efficiency gains, establishing hybrid binarization as a practical and scalable approach for efficient neural video representation.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 153864-153883
TinyNeRV: Compact Neural Video Representations via Capacity Scaling, Distillation, and Low-Precision Inference
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 153864-153883
Open AccessTinyNeRV: Compact Neural Video Representations via Capacity Scaling, Distillation, and Low-Precision Inference
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 97418-97431
LRConv-NeRV: Low-Rank Convolution for Efficient Neural Video Compression
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 97418-97431
Open AccessLRConv-NeRV: Low-Rank Convolution for Efficient Neural Video Compression
- 1 July 2025
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 84 (37) , 46457-46478
Dynamic video summarization using handcrafted features to complement publicly available datasets
- 1 July 2025
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 84 (37) , 46457-46478
Dynamic video summarization using handcrafted features to complement publicly available datasets
- Open Access
- 25 April 2025
- Published by MDPI AG in Information
- Vol. 16 (5) , 348
Edge-Optimized Deep Learning Architectures for Classification of Agricultural Insects with Mobile Deployment
The deployment of machine learning models on mobile platforms has ushered in a new era of innovation across diverse sectors, including agriculture, where such applications hold immense promise for empowering farmers with cutting-edge technologies. In this context, the threat posed by insects to crop yields during harvest has escalated, fueled by factors such as evolution and climate change-induced shifts in insect behavior. To address this challenge, smart insect monitoring systems and detection models have emerged as crucial tools for farmers and IoT-based systems, enabling interventions to safeguard crops. The primary contribution of this study lies in its systematic investigation of model optimization techniques for edge deployment, including Post-Training Quantization, Quantization-Aware Training, and Data Representative Quantization. As such, we address the crucial need for efficient, on-site pest detection tools in agricultural settings. We provide a detailed analysis of the trade-offs between model size, inference speed, and accuracy across different optimization approaches, ensuring practical applicability in resource-constrained farming environments. Our study explores various methodologies for model development, including the utilization of Mobile-ViT and EfficientNet architectures, coupled with transfer learning and fine-tuning techniques. Using the Dangerous Farm Insects Dataset, we achieve an accuracy of 82.6% and 77.8% on validation and test datasets, respectively, showcasing the efficacy of our approach. Furthermore, we investigate quantization techniques to optimize model performance for on-device inference, ensuring seamless deployment on mobile devices and other edge devices without compromising accuracy. The best quantized model, produced through Post-Training Quantization, was able to maintain a classification accuracy of 77.8% while significantly reducing the model size from 33 MB to 9.6 MB. To validate the generalizability of our solution, we extended our experiments to the larger IP102 dataset. The quantized model produced using Post-Training Quantization was able to maintain a classification accuracy of 59.6% while also reducing the model size from 33 MB to 9.6 MB, thus demonstrating that our solution maintains a competitive performance across a broader range of insect classes.
- 25 April 2025
- Published by MDPI AG in Information
- Vol. 16 (5) , 348
Open AccessEdge-Optimized Deep Learning Architectures for Classification of Agricultural Insects with Mobile Deployment
The deployment of machine learning models on mobile platforms has ushered in a new era of innovation across diverse sectors, including agriculture, where such applications hold immense promise for empowering farmers with cutting-edge technologies. In this context, the threat posed by insects to crop yields during harvest has escalated, fueled by factors such as evolution and climate change-induced shifts in insect behavior. To address this challenge, smart insect monitoring systems and detection models have emerged as crucial tools for farmers and IoT-based systems, enabling interventions to safeguard crops. The primary contribution of this study lies in its systematic investigation of model optimization techniques for edge deployment, including Post-Training Quantization, Quantization-Aware Training, and Data Representative Quantization. As such, we address the crucial need for efficient, on-site pest detection tools in agricultural settings. We provide a detailed analysis of the trade-offs between model size, inference speed, and accuracy across different optimization approaches, ensuring practical applicability in resource-constrained farming environments. Our study explores various methodologies for model development, including the utilization of Mobile-ViT and EfficientNet architectures, coupled with transfer learning and fine-tuning techniques. Using the Dangerous Farm Insects Dataset, we achieve an accuracy of 82.6% and 77.8% on validation and test datasets, respectively, showcasing the efficacy of our approach. Furthermore, we investigate quantization techniques to optimize model performance for on-device inference, ensuring seamless deployment on mobile devices and other edge devices without compromising accuracy. The best quantized model, produced through Post-Training Quantization, was able to maintain a classification accuracy of 77.8% while significantly reducing the model size from 33 MB to 9.6 MB. To validate the generalizability of our solution, we extended our experiments to the larger IP102 dataset. The quantized model produced using Post-Training Quantization was able to maintain a classification accuracy of 59.6% while also reducing the model size from 33 MB to 9.6 MB, thus demonstrating that our solution maintains a competitive performance across a broader range of insect classes.
- Open Access
Content-Symmetrical Multidimensional Transpose of Image Sequences for the High Efficiency Video Coding (HEVC) All-Intra Configuration
Enhancing the quality of video coding whilst maintaining compliance with the syntax of video coding standards is challenging. In the literature, many solutions have been proposed that apply mainly to two-pass encoding, bitrate control algorithms, and enhancements of locally decoded images in the motion-compensation loop. This work proposes a pre- and post-coding solution using the content-symmetrical multidimensional transpose of raw video sequences. The content-symmetrical multidimensional transpose results in images composed of slices of the temporal domain whilst preserving the video content. Such slices have higher spatial homogeneity at the expense of reducing the temporal resemblance. As such, an all-intra configuration is an excellent choice for compressing such images. Prior to displaying the decoded images, a content-symmetrical multidimensional transpose is applied again to restore the original form of the input images. Moreover, we propose a lightweight two-pass encoding solution in which we apply systematic temporal subsampling on the multidimensional transposed image sequences prior to the first-pass encoding. This noticeably reduces the complexity of the encoding process of the first pass and gives an indication as to whether or not the proposed solution is suitable for the video sequence at hand. Using the HEVC video codec, the experimental results revealed that the proposed solution results in a lower percentage of coding unit splits in comparison to regular HEVC coding without the multidimensional transpose of image sequences. This finding supports the claim of there being increasing spatial coherence as a result of the proposed solution. Additionally, using four quantization parameters, and in comparison to regular HEVC encoding, the resulting BD rate is −15.12%, which indicates a noticeable bitrate reduction. The BD-PSNR, on the other hand, was 1.62 dB, indicating an enhancement in the quality of the decoded images. Despite all of these benefits, the proposed solution has limitations, which are also discussed in the paper.
Open AccessContent-Symmetrical Multidimensional Transpose of Image Sequences for the High Efficiency Video Coding (HEVC) All-Intra Configuration
Enhancing the quality of video coding whilst maintaining compliance with the syntax of video coding standards is challenging. In the literature, many solutions have been proposed that apply mainly to two-pass encoding, bitrate control algorithms, and enhancements of locally decoded images in the motion-compensation loop. This work proposes a pre- and post-coding solution using the content-symmetrical multidimensional transpose of raw video sequences. The content-symmetrical multidimensional transpose results in images composed of slices of the temporal domain whilst preserving the video content. Such slices have higher spatial homogeneity at the expense of reducing the temporal resemblance. As such, an all-intra configuration is an excellent choice for compressing such images. Prior to displaying the decoded images, a content-symmetrical multidimensional transpose is applied again to restore the original form of the input images. Moreover, we propose a lightweight two-pass encoding solution in which we apply systematic temporal subsampling on the multidimensional transposed image sequences prior to the first-pass encoding. This noticeably reduces the complexity of the encoding process of the first pass and gives an indication as to whether or not the proposed solution is suitable for the video sequence at hand. Using the HEVC video codec, the experimental results revealed that the proposed solution results in a lower percentage of coding unit splits in comparison to regular HEVC coding without the multidimensional transpose of image sequences. This finding supports the claim of there being increasing spatial coherence as a result of the proposed solution. Additionally, using four quantization parameters, and in comparison to regular HEVC encoding, the resulting BD rate is −15.12%, which indicates a noticeable bitrate reduction. The BD-PSNR, on the other hand, was 1.62 dB, indicating an enhancement in the quality of the decoded images. Despite all of these benefits, the proposed solution has limitations, which are also discussed in the paper.