Lightweight Model Research Articles

聚焦可变形注意力的马铃薯芽苗精确无损检测方法

Accurate potato sprout detection is the key to automatic seed potato cutting, which is important for potato quality and yield. In this paper, a lightweight DAS-YOLOv8 model is proposed for the potato sprout detection task. By embedding DAS deformable attention in the feature extraction network and the feature fusion network, the global feature context can be efficiently represented and the attention increased to the relevant pixel image region; then, the C2f_Atten module fusing Shuffle attention is designed based on the C2f module to satisfy the attention to the key feature information of the high-level abstract semantics of the feature extraction network. At the same time, the ghost convolution is introduced to improve the C2f module and convolutional module to realize the decomposition of the redundant features to extract the key features. Verified on the collected potato sprout image data set, the average accuracy of the proposed DAS-YOLOv8 model is 94.25%, and the calculation amount is only 7.66 G. Compared with the YOLOv8n model, the accuracy is 2.13% higher, and the average accuracy is 1.55% higher. In comparison to advanced state-of-the-art (SOTA) target detection algorithms, the method in this paper offers a better balance between comprehensive performance and lightweight model design. The improved and optimized DAS-YOLOv8 model can realize the effective detection of potato sprouts, meet the requirements of real-time processing, and can provide theoretical support for the non-destructive detection of sprouts in automatic seed potato cutting.

S-YOLOv5: A Lightweight Model for Detecting Objects Thrown from Tall Buildings in Communities

Objects thrown from tall buildings in communities are characterized by their small size, inconspicuous features, and high speed. Existing algorithms for detecting such objects face challenges, including excessive parameters, overly complex models that are difficult to implement, and insufficient detection accuracy. This study proposes a lightweight detection model for objects thrown from tall buildings in communities, named S-YOLOv5, to address these issues. The model is based on the YOLOv5 algorithm, and a lightweight convolutional neural network, Enhanced ShuffleNet (ESNet), is chosen as the backbone network to extract image features. On this basis, the initial stage of the backbone network is enhanced and the simplified attention module (SimAM) attention mechanism is added to utilize the rich position information and contour information in the shallow feature map to improve the detection of small targets. For feature fusion, the sparsely connected Path Aggregation Network (SCPANet) module is designed to use sparsely connected convolution (SCConv) instead of the regular convolution of the Path Aggregation Network (PANet) to fuse features efficiently. In addition, the model uses the normalized Wasserstein distance (NWD) loss function to reduce the sensitivity of positional bias. The accuracy of the model is further improved. Test results from the self-built objects thrown from tall buildings dataset show that S-YOLOv5 can detect objects thrown from tall buildings quickly and accurately, with an accuracy of 90.2% and a detection rate of 34.1 Fps/s. Compared with the original YOLOv5 model, the parameters are reduced by 87.3%, and the accuracy and rate are improved by 0.8% and 63%, respectively.

A Study for Design of Lightweight Dense Connection Network on Hyperspectral Image Classification

The characteristics of hyperspectral remote sensing images such as inconspicuous feature representativeness, single feature level, and complex information content, can lead to unstable classification results. We propose a lightweight dense network model that injects channel attention in the form of dense connections between network layers (DSE-DN) for the classification of hyperspectral images. In the DSE-DN network, principal component analysis (PCA) is applied to reduce redundancy in the hyperspectral images. Subsequently, a densely connected network is constructed, incorporating channel attention mechanisms through dense connections to enhance the analysis of spectral image features. Finally, the processed hyperspectral images are classified using a fully interconnected layer. We assess two classical hyperspectral datasets and construct 2DCNN, 3DCNN, ResNet, and the network that injects channel attention layer by layer to compare with DSE-DN. The experimental results indicate the utility of the DSE-DN network in hyperspectral image classification and its superiority over other networks.

YOLOv8s-CGF: a lightweight model for wheat ear Fusarium head blight detection.

Fusarium head blight (FHB) is a destructive disease that affects wheat production. Detecting FHB accurately and rapidly is crucial for improving wheat yield. Traditional models are difficult to apply to mobile devices due to large parameters, high computation, and resource requirements. Therefore, this article proposes a lightweight detection method based on an improved YOLOv8s to facilitate the rapid deployment of the model on mobile terminals and improve the detection efficiency of wheat FHB. The proposed method introduced a C-FasterNet module, which replaced the C2f module in the backbone network. It helps reduce the number of parameters and the computational volume of the model. Additionally, the Conv in the backbone network is replaced with GhostConv, further reducing parameters and computation without significantly affecting detection accuracy. Thirdly, the introduction of the Focal CIoU loss function reduces the impact of sample imbalance on the detection results and accelerates the model convergence. Lastly, the large target detection head was removed from the model for lightweight. The experimental results show that the size of the improved model (YOLOv8s-CGF) is only 11.7 M, which accounts for 52.0% of the original model (YOLOv8s). The number of parameters is only 5.7 × 106 M, equivalent to 51.4% of the original model. The computational volume is only 21.1 GFLOPs, representing 74.3% of the original model. Moreover, the mean average precision (mAP@0.5) of the model is 99.492%, which is 0.003% higher than the original model, and the mAP@0.5:0.95 is 0.269% higher than the original model. Compared to other YOLO models, the improved lightweight model not only achieved the highest detection precision but also significantly reduced the number of parameters and model size. This provides a valuable reference for FHB detection in wheat ears and deployment on mobile terminals in field environments.

Mammo-Light: A lightweight convolutional neural network for diagnosing breast cancer from mammography images

People of all countries, developed and developing alike endure cancer-related fatal diseases. The rate of breast cancer in females is increasing daily, partly due to ignorance and misdiagnosis in the early stages. Diagnosis of breast cancer accurately during its earlier stages of development can result in proper initial treatment for breast cancer. Artificial intelligence can aid in the acceleration and automation of breast cancer detection. Deep learning is decisive in effectively recognizing and classifying cancer on large datasets of medical images. In this paper, we propose a novel computer-aided classification approach, Mammo-Light for breast cancer prediction. Preprocessing strategies have been utilized to eradicate the noise and enhance mammogram lesions. Photometric augmentation techniques adapted to the preprocessed classes to balance and increase the size of the dataset. After that, a lightweight yet intuitive convolutional neural network is applied to classify breast cancer on the publicly available dataset CBIS-DDSM. For further validation of the proposed approach, we have used the MIAS dataset. Mammo-Light attained a 99.17% and 98.42% test accuracy respectively for CBIS-DDSM and MIAS datasets and outperformed state-of-the-art methods in terms of accuracy and other metrics. Due to being the lightweight model, Mammo-Light performs exceptionally well with fewer parameters and computational time, which can potentially contribute to the field of breast cancer early diagnosis and enable fast treatment.

Detection of Rice Leaf Diseases Based on Improved YOLOv8n

Rice plays an important role in human food chain, but it is easily affected by related diseases in its growth, which seriously affects rice yield. At present, there are many methods to detect rice leaf diseases, but there are still problems such as large model and poor detection effect. In order to solve the limitations of traditional manual and deep learning-based models, a lightweight model based on improved YOLOv8n is proposed. Firstly, SPPD (Spatial Pyramid Pooling and Dilated) structure with different void ratios composed of GELU activation function was added to the Backbone network to increase the receptive field of the network, and Coordinate attention (CA) was combined to help the model pay attention to the characteristics of rice diseases and improve the detection accuracy. Finally, XSEPConv (Extremely Separated Convolution) is used to reduce the parameters and improve the efficiency of the model. The model was trained and tested on the self-built rice leaf disease image data set, and its mAP @0.5 reached 89.3% and FPS reached 217. The proposed SCX-YOLOv8n model is a lightweight, efficient and high-performance rice leaf disease detection model, and compared with other mainstream models, it also has certain advantages in accuracy and recall rate, which can provide an accurate and accurate rice leaf disease detection and other related fields.

Lightweight detection network for bridge defects based on model pruning and knowledge distillation

Deep learning techniques improve the detection performance of objects by expanding the depth of network and increasing the complexity of model structure, however, this makes it difficult to deploy as an application on embedded devices. In this study a lightweight model was put forward to better identify bridge defects combining with two advanced feature fusion modules to increase the multi-scale information perception and extraction capability of the network. Sparse training and slimming pruning were used to remove redundant parameters. Then, fine-tuning and knowledge distillation (KD) enabled the network to recover its detection accuracy and robustness. Furthermore, the unsupervised learning CycleGAN was used to enhance the generalization performance of the dataset and improve the detection accuracy. The experimental results show that compared with the original model, the precision rate, mAP@ 0.5 and detection speed (FPS) is increased by 8.07%, 6.15% and 85.06%, respectively. Moreover, the model volume of proposed network is 43.04% of the original ones. The results indicates that the proposed network can be deployed on devices with limited storage to achieve accurate and real-time detection of bridge defects. These findings represent a major step forward in the automation of detection and quantification of structural defects.

An efficient intrusion detection model based on convolutional spiking neural network

Many intrusion detection techniques have been developed to ensure that the target system can function properly under the established rules. With the booming Internet of Things (IoT) applications, the resource-constrained nature of its devices makes it urgent to explore lightweight and high-performance intrusion detection models. Recent years have seen a particularly active application of deep learning (DL) techniques. The spiking neural network (SNN), a type of artificial intelligence that is associated with sparse computations and inherent temporal dynamics, has been viewed as a potential candidate for the next generation of DL. It should be noted, however, that current research into SNNs has largely focused on scenarios where limited computational resources and insufficient power sources are not considered. Consequently, even state-of-the-art SNN solutions tend to be inefficient. In this paper, a lightweight and effective detection model is proposed. With the help of rational algorithm design, the model integrates the advantages of SNNs as well as convolutional neural networks (CNNs). In addition to reducing resource usage, it maintains a high level of classification accuracy. The proposed model was evaluated against some current state-of-the-art models using a comprehensive set of metrics. Based on the experimental results, the model demonstrated improved adaptability to environments with limited computational resources and energy sources.

Comparative analysis of strategies of knowledge distillation on BERT for text matching

Large language model is a highly effective and promising language model technology that can improve the performance and robustness of natural language processing tasks. As a representative work, Bidirectional Encoder Representations from Transformers (BERT) has excellent performance in various natural language processing tasks. This model is pre-trained on large-scale language dataset and has gained attention from all walks of life since its introduction. However, its huge number of participants and scale make its performance in mobile very limited. As an effective technique to compress neural network, knowledge distillation can obtain a lightweight model with smaller parameters without losing too much model performance. Therefore, distillation of BERT models has been started, aiming at obtaining lightweight BERT models. In this paper, we will introduce several common BERT distillation models and analyse their model architecture, distillation process, and finally the compression efficiency and model effectiveness. It is concluded that the process of increasing recompression efficiency is often accompanied by decreasing model effectiveness.

FlexKBQA: A Flexible LLM-Powered Framework for Few-Shot Knowledge Base Question Answering

Knowledge base question answering (KBQA) is a critical yet challenging task due to the vast number of entities within knowledge bases and the diversity of natural language questions posed by users. Unfortunately, the performance of most KBQA models tends to decline significantly in real-world scenarios where high-quality annotated data is insufficient. To mitigate the burden associated with manual annotation, we introduce FlexKBQA by utilizing Large Language Models (LLMs) as program translators for addressing the challenges inherent in the few-shot KBQA task. Specifically, FlexKBQA leverages automated algorithms to sample diverse programs, such as SPARQL queries, from the knowledge base, which are subsequently converted into natural language questions via LLMs. This synthetic dataset facilitates training a specialized lightweight model for the KB. Additionally, to reduce the barriers of distribution shift between synthetic data and real user questions, FlexKBQA introduces an executionguided self-training method to iterative leverage unlabeled user questions. Furthermore, we explore harnessing the inherent reasoning capability of LLMs to enhance the entire framework. Consequently, FlexKBQA delivers substantial flexibility, encompassing data annotation, deployment, and being domain agnostic. Through extensive experiments on GrailQA, WebQSP, and KQA Pro, we observe that under the few-shot even the more challenging zero-shot scenarios, FlexKBQA achieves impressive results with a few annotations, surpassing all previous baselines and even approaching the performance of supervised models, achieving a remarkable 93% performance relative to the fully-supervised models. We posit that FlexKBQA represents a significant advancement towards exploring better integration of large and lightweight models. Code is available at https://github.com/leezythu/FlexKBQA.

LaMAR: Laplacian Pyramid for Multimodal Adaptive Super Resolution (Student Abstract)

Recent advances in image-to-image translation involve the integration of non-visual imagery in deep models. Non-visual sensors, although more costly, often produce low-resolution images. To combat this, methods using RGB images to enhance the resolution of these modalities have been introduced. Fusing these modalities to achieve high-resolution results demands models with millions of parameters and extended inference times. We present LaMAR, a lightweight model. It employs Laplacian image pyramids combined with a low-resolution thermal image for Guided Thermal Super Resolution. By decomposing the RGB image into a Laplacian pyramid, LaMAR preserves image details and avoids high-resolution feature map computations, ensuring efficiency. With faster inference times and fewer parameters, our model demonstrates state-of-the-art results.

FRIH: Fine-Grained Region-Aware Image Harmonization

Image harmonization aims to generate a more realistic appearance of foreground and background for a composite image. All the existing methods perform the same harmonization process for the whole foreground. However, the implanted foreground always contains different appearance patterns. Existing solutions ignore the difference of each color block and lose some specific details. Therefore, we propose a novel global-local two stages framework for Fine-grained Region-aware Image Harmonization (FRIH). In the first stage, the whole input foreground mask is used to make a global coarse-grained harmonization. In the second stage, we adaptively cluster the input foreground mask into several submasks. Each submask and the coarsely adjusted image are concatenated respectively and fed into a lightweight cascaded module, refining the global harmonization result. Moreover, we further design a fusion prediction module to generate the final result, utilizing the different degrees of harmonization results comprehensively. Without bells and whistles, our FRIH achieves a competitive performance on iHarmony4 dataset with a lightweight model.

AesFA: An Aesthetic Feature-Aware Arbitrary Neural Style Transfer

Neural style transfer (NST) has evolved significantly in recent years. Yet, despite its rapid progress and advancement, existing NST methods either struggle to transfer aesthetic information from a style effectively or suffer from high computational costs and inefficiencies in feature disentanglement due to using pre-trained models. This work proposes a lightweight but effective model, AesFA---Aesthetic Feature-Aware NST. The primary idea is to decompose the image via its frequencies to better disentangle aesthetic styles from the reference image while training the entire model in an end-to-end manner to exclude pre-trained models at inference completely. To improve the network's ability to extract more distinct representations and further enhance the stylization quality, this work introduces a new aesthetic feature: contrastive loss. Extensive experiments and ablations show the approach not only outperforms recent NST methods in terms of stylization quality, but it also achieves faster inference. Codes are available at https://github.com/Sooyyoungg/AesFA.

Behavioral Recognition of Skeletal Data Based on Targeted Dual Fusion Strategy

The deployment of multi-stream fusion strategy on behavioral recognition from skeletal data can extract complementary features from different information streams and improve the recognition accuracy, but suffers from high model complexity and a large number of parameters. Besides, existing multi-stream methods using a fixed adjacency matrix homogenizes the model’s discrimination process across diverse actions, causing reduction of the actual lift for the multi-stream model. Finally, attention mechanisms are commonly applied to the multi-dimensional features, including spatial, temporal and channel dimensions. But their attention scores are typically fused in a concatenated manner, leading to the ignorance of the interrelation between joints in complex actions. To alleviate these issues, the Front-Rear dual Fusion Graph Convolutional Network (FRF-GCN) is proposed to provide a lightweight model based on skeletal data. Targeted adjacency matrices are also designed for different front fusion streams, allowing the model to focus on actions of varying magnitudes. Simultaneously, the mechanism of Spatial-Temporal-Channel Parallel Attention (STC-P), which processes attention in parallel and places greater emphasis on useful information, is proposed to further improve model’s performance. FRF-GCN demonstrates significant competitiveness compared to the current state-of-the-art methods on the NTU RGB+D, NTU RGB+D 120 and Kinetics-Skeleton 400 datasets. Our code is available at: https://github.com/sunbeam-kkt/FRF-GCN-master.

Efficient Speech Detection in Environmental Audio Using Acoustic Recognition and Knowledge Distillation.

The ongoing biodiversity crisis, driven by factors such as land-use change and global warming, emphasizes the need for effective ecological monitoring methods. Acoustic monitoring of biodiversity has emerged as an important monitoring tool. Detecting human voices in soundscape monitoring projects is useful both for analyzing human disturbance and for privacy filtering. Despite significant strides in deep learning in recent years, the deployment of large neural networks on compact devices poses challenges due to memory and latency constraints. Our approach focuses on leveraging knowledge distillation techniques to design efficient, lightweight student models for speech detection in bioacoustics. In particular, we employed the MobileNetV3-Small-Pi model to create compact yet effective student architectures to compare against the larger EcoVAD teacher model, a well-regarded voice detection architecture in eco-acoustic monitoring. The comparative analysis included examining various configurations of the MobileNetV3-Small-Pi-derived student models to identify optimal performance. Additionally, a thorough evaluation of different distillation techniques was conducted to ascertain the most effective method for model selection. Our findings revealed that the distilled models exhibited comparable performance to the EcoVAD teacher model, indicating a promising approach to overcoming computational barriers for real-time ecological monitoring.

A novel lightweight model HGCA-YOLO: Application to recognition of invisible spears for white asparagus robotic harvesting

Timely harvesting white asparagus depend on the detection of two forms: the emerging unearthed spear tips and the soil leaks raised by earthed spears. Accurate detection of the small spear tips and the invisible spears in the complex backgrounds of field ridges remains a challenge, especially the soil leaks with pattern similarity to the drought-induced cracks. In this paper, a novel lightweight model named HGCA-YOLO (Hyperparameter evolution-Ghost module-Coordinate Attention mechanism You Only Look Once algorithm) is proposed for accurate detection of two forms of targets. Firstly, the baseline network is determined by adopting the hyperparameter evolution to converge the network faster and obtain better parameters for spear detection. Then, the Ghost module and coordinate attention mechanism are introduced in the baseline network to decrease the complexity of the model as well as to enhance the sensitivity to the target location. In addition, the TTA (Test Time Augmentation) is introduced to the network inference to handle the targets in strongly varying environment. Finally, a dataset covered spear tips and soil leaks acquired from natural ridge is constructed, and the detection test and field experiments are conducted. The experimental results show that the accuracy of the proposed method achieved mAP 0.952 and mF1 0.924. Compared to the baseline network, this method reduced the parameters, GFLOPs and model size by 46.2%, 48.4% and 44.7%. In particular, the success rate of detection achieved 87% in the field test. This lightweight model effectively extracted the features of the cracking pattern of invisible asparagus and the small spear tips, which improved the accuracy of spear detection for selective robotic harvesting of white asparagus.

SCSONet: spatial-channel synergistic optimization net for skin lesion segmentation

In the field of computer-assisted medical diagnosis, developing medical image segmentation models that are both accurate and capable of real-time operation under limited computational resources is crucial. Particularly for skin disease image segmentation, the construction of such lightweight models must balance computational cost and segmentation efficiency, especially in environments with limited computing power, memory, and storage. This study proposes a new lightweight network designed specifically for skin disease image segmentation, aimed at significantly reducing the number of parameters and floating-point operations while ensuring segmentation performance. The proposed ConvStem module, with full-dimensional attention, learns complementary attention weights across all four dimensions of the convolution kernel, effectively enhancing the recognition of irregularly shaped lesion areas, reducing the model’s parameter count and computational burden, thus promoting model lightweighting and performance improvement. The SCF Block reduces feature redundancy through spatial and channel feature fusion, significantly lowering parameter count while improving segmentation results. This paper validates the effectiveness and robustness of the proposed SCSONet on two public skin lesion segmentation datasets, demonstrating its low computational resource requirements. https://github.com/Haoyu1Chen/SCSONet.

Enhancing Sun-Dried Kelp Detection: Introducing K-YOLO, a Lightweight Model with Improved Precision and Recall.

Kelp, often referred to as a "sea vegetable", holds substantial economic significance. Currently, the drying process for kelp in China primarily relies on outdoor sun-drying methods. Detecting kelp in the field presents challenges arising from issues such as overlapping and obstruction. To address these challenges, this study introduces a lightweight model, K-YOLOv5, specifically designed for the precise detection of sun-dried kelp. YOLOv5-n serves as the base model, with several enhancements implemented in this study: the addition of a detection head incorporating an upsampling layer and a convolution module to improve the recognition of small objects; the integration of an enhanced I-CBAM attention mechanism, focusing on key features to enhance the detection accuracy; the replacement of the CBS module in the neck network with GSConv to reduce the computational burden and accelerate the inference speed; and the optimization of the IoU algorithm to improve the identification of overlapping kelp. Utilizing drone-captured images of sun-dried kelp, a dataset comprising 2190 images is curated. Validation on this self-constructed dataset indicates that the improved K-YOLOv5 model significantly enhances the detection accuracy, achieving 88% precision and 78.4% recall. These values represent 6.8% and 8.6% improvements over the original model, respectively, meeting the requirements for the real-time recognition of sun-dried kelp.

FLP: Factor lattice pattern-based automated detection of Parkinson's disease and specific language impairment using recorded speech

BackgroundTimely detection of neurodevelopmental and neurological conditions is crucial for early intervention. Specific Language Impairment (SLI) in children and Parkinson's disease (PD) manifests in speech disturbances that may be exploited for diagnostic screening using recorded speech signals. We were motivated to develop an accurate yet computationally lightweight model for speech-based detection of SLI and PD, employing novel feature engineering techniques to mimic the adaptable dynamic weight assignment network capability of deep learning architectures. Materials and methodsIn this research, we have introduced an advanced feature engineering model incorporating a novel feature extraction function, the Factor Lattice Pattern (FLP), which is a quantum-inspired method and uses a superposition-like mechanism, making it dynamic in nature. The FLP encompasses eight distinct patterns, from which the most appropriate pattern was discerned based on the data structure. Through the implementation of the FLP, we automatically extracted signal-specific textural features. Additionally, we developed a new feature engineering model to assess the efficacy of the FLP. This model is self-organizing, producing nine potential results and subsequently choosing the optimal one. Our speech classification framework consists of (1) feature extraction using the proposed FLP and a statistical feature extractor; (2) feature selection employing iterative neighborhood component analysis and an intersection-based feature selector; (3) classification via support vector machine and k-nearest neighbors; and (4) outcome determination using combinational majority voting to select the most favorable results. ResultsTo validate the classification capabilities of our proposed feature engineering model, designed to automatically detect PD and SLI, we employed three speech datasets of PD and SLI patients. Our presented FLP-centric model achieved classification accuracy of more than 95% and 99.79% for all PD and SLI datasets, respectively. ConclusionsOur results indicate that the proposed model is an accurate alternative to deep learning models in classifying neurological conditions using speech signals.

A knowledge distillation based deep learning framework for cropped images detection in spatial domain

Cropping an image is a common image editing technique that aims to find viewpoints with suitable image composition. It is also a frequently used post-processing technique to reduce the evidence of tampering in an image. Detecting cropped images poses a significant challenge in the field of digital image forensics, as the distortions introduced by image cropping are often imperceptible to the human eye. Although deep neural networks achieve state-of-the-art performance, due to their ability to encode large-scale data and handle billions of model parameters. However, due to their high computational complexity and substantial storage requirements, it is difficult to deploy these large deep learning models on resource-constrained devices such as mobile phones and embedded systems. To address this issue, we propose a lightweight deep learning framework for cropping detection in spatial domain, based on knowledge distillation. Initially, we constructed four datasets containing a total of 60,000 images cropped using various tools. We then used Efficient-Net-B0, pre-trained on ImageNet with significant surgical adjustments, as the teacher model, which makes it more robust and faster to converge in this downstream task. The model was trained on 20,000 cropped and uncropped images from our own dataset, and we then applied its knowledge to a more compact model called the student model. Finally, we selected the best-performing lightweight model as the final prediction model, with a testing accuracy of 98.44% on the test dataset, which outperforms other methods. Extensive experiments demonstrate that our proposed model, distilled from Efficient-Net-B0, achieves state-of-the-art performance in terms of detection accuracy, training parameters, and FLOPs, outperforming existing methods in detecting cropped images.