We study the challenging task of neural network quantization without end-to-end retraining, called Post-training Quantization (PTQ). PTQ usually requires a small subset training data but produces less powerful quantized models than Quantization-Aware Training (QAT). In this work, we propose novel framework, dubbed BRECQ, which pushes limits bitwidth in down to INT2 for first time. BRECQ leverages basic building blocks networks and reconstructs them one-by-one. comprehensive theoretical...

Quantization has emerged as one of the most prevalent approaches to compress and accelerate neural networks. Recently, data-free quantization been widely studied a practical promising solution. It synthesizes data for calibrating quantized model according batch normalization (BN) statistics FP32 ones significantly relieves heavy dependency on real training in traditional methods. Unfortunately, we find that practice, synthetic identically constrained by BN suffers serious homogenization at...

Recently, post-training quantization (PTQ) has driven much attention to produce efficient neural networks without long-time retraining. Despite its low cost, current PTQ works tend fail under the extremely low-bit setting. In this study, we pioneeringly confirm that properly incorporating activation into reconstruction benefits final accuracy. To deeply understand inherent reason, a theoretical framework is established, indicating flatness of optimized model on calibration and test data...

Unsupervised domain adaptation (UDA) in medical image segmentation aims to improve the generalization of deep models by alleviating gaps caused inconsistency across equipment, imaging protocols, and patient conditions. However, existing UDA works remain insufficiently explored present great limitations: (i) Exhibit cumbersome designs that prioritize aligning statistical metrics distributions, which limits model's flexibility while also overlooking potential knowledge embedded unlabeled data;...

Generative dataset expansion methods can effectively alleviate the scarcity of data in dermoscopic image segmentation but commonly employ a two-stage synthesis strategy that contains additional learnable components and complex design, which results high computational resource costs. Diffusion models utilizing self-conditioning have shown strong potential for efficiently reusing priors pipeline without relying on excessively complicated conditioning designs. Inspired by this, we propose...

Quantization Neural Networks (QNN) have attracted a lot of attention due to their high efficiency. To enhance the quantization accuracy, prior works mainly focus on designing advanced algorithms but still fail achieve satisfactory results under extremely low-bit case. In this work, we take an architecture perspective investigate potential high-performance QNN. Therefore, propose combine Network Architecture Search methods with enjoy merits two sides. However, naive combination inevitably...

To address the challenges of balancing accuracy and speed, as well parameters FLOPs in current insulator defect detection, we propose an enhanced detection algorithm, ML-YOLOv5, based on YOLOv5 network. The backbone module incorporates depthwise separable convolution, feature fusion C3 is replaced with improved C2f_DG module. Furthermore, enhance pyramid network (MFPN) employ knowledge distillation using YOLOv5m teacher model. Experimental results demonstrate that this approach achieved a...

To deploy deep neural networks on resource-limited devices, quantization has been widely explored. In this work, we study the extremely low-bit which have tremendous speed-up, memory saving with quantized activation and weights. We first bring up three omitted issues in networks: squashing range of values; gradient vanishing during backpropagation unexploited hardware acceleration ternary networks. By reparameterizing weights vector full precision scale offset for fixed vector, decouple...

User data confidentiality protection is becoming a rising challenge in the present deep learning research. Without access to data, conventional data-driven model compression faces higher risk of performance degradation. Recently, some works propose generate images from specific pretrained serve as training data. However, inversion process only utilizes biased feature statistics stored one and low-dimension high-dimension. As consequence, it inevitably encounters difficulties generalizability...

Model quantization has emerged as an indispensable technique to accelerate deep learning inference. While researchers continue push the frontier of algorithms, existing work is often unreproducible and undeployable. This because do not choose consistent training pipelines ignore requirements for hardware deployments. In this work, we propose Quantization Benchmark (MQBench), a first attempt evaluate, analyze, benchmark reproducibility deployability model algorithms. We multiple different...

In this study, we explore Human Activity Recognition (HAR), a task that aims to predict individuals' daily activities utilizing time series data obtained from wearable sensors for health-related applications. Although recent research has predominantly employed end-to-end Artificial Neural Networks (ANNs) feature extraction and classification in HAR, these approaches impose substantial computational load on devices exhibit limitations temporal due their activation functions. To address...

Video frame interpolation aims to generate intermediate frames in a video showcase finer details. However, most methods are only trained and tested on low-resolution datasets, lacking research 4K problems. This limitation makes it challenging handle high-frame-rate processing real-world scenarios. In this paper, we propose dataset at 120 fps, named UHD4K120FPS, which contains large motion. We also novel framework for solving the task, based multi-scale pyramid network structure. introduce...

The environmental perception of autonomous vehicles in normal conditions have achieved considerable success the past decade. However, various unfavourable such as fog, low-light, and motion blur will degrade image quality pose tremendous threats to safety driving. That is, when applied degraded images, state-of-the-art visual models often suffer performance decline due feature content loss artifact interference caused by statistical structural properties disruption captured images. To...

The environmental perception of autonomous vehicles in normal conditions have achieved considerable success the past decade. However, various unfavourable such as fog, low-light, and motion blur will degrade image quality pose tremendous threats to safety driving. That is, when applied degraded images, state-of-the-art visual models often suffer performance decline due feature content loss artifact interference caused by statistical structural properties disruption captured images. To...

This paper proposes a semi-fragile digital watermarking scheme for image tamper localization and self-recovery. Firstly, the authentication watermark is generated tampered area localization. After that, content recovery, recovery calculated from high frequency band with redundancy free. Then, two type of watermarks, DWT-based embedding method makes our proposed tolerable against friendly manipulations. Even some parts watermarked are corresponding embedded information missing, we can still...