Model Size Reduction Research Articles

TEC-CNN: Towards Efficient Compressing Convolutional Neural Nets with Low-rank Tensor Decomposition

Most state-of-the-art convolutional neural networks (CNNs) are characterised by excessive parameterisation, leading to a high computational burden. Tensor decomposition has emerged as a model reduction technique for compressing deep neural networks. Previous approaches have predominantly relied on either Tucker decomposition or Canonical Polyadic (CP) decomposition for CNNs. However, CP decomposition exhibits exceptional compression capabilities in comparison to Tucker decomposition, which results in a more pronounced accuracy loss. This paper introduces an efficient model compression method, termed TEC-CNN, designed to achieve significant compression while preserving accuracy levels comparable to those of the original models. In TEC-CNN, convolutional layers are identified to obtain convolutional kernels by analysing given models under the principles of low-rank tensor decomposition, and then, calculating the ranks of convolutional kernels. Furthermore, an efficient decomposition schema for the convolutional kernel is proposed with approximate kernel tensor for reducing parameters. Additionally, a novel format of a convolutional sequence is presented and constructed with a reduced number of parameters to replace the original convolutional layers. Finally, the effectiveness of TEC-CNN is assessed across a range of computer vision tasks. For instance, in CIFAR-100 classification, ResNet18 is compressed to 4.1 MB, while Unext, when applied to image segmentation using the International Skin Imaging Collaboration (ISIC) dataset, is reduced to 3.419 MB. When employed for fire object detection with Yolov7, TEC-CNN achieves a model size reduction of 71.6 MB. Comprehensive experimental results underscore that our approach achieves significant model compression while preserving model performance.

Adaptive Transformer-Based Multi-Modal Image Fusion for Real-Time Medical Diagnosis and Object Detection

In recent years, medical diagnosis and object detection have been significantly enhanced by the integration of multi-modal image fusion techniques. This study proposes an Adaptive Transformer-Based Multi-Modal Image Fusion (AT-MMIF) framework designed for real-time medical diagnosis and object detection. The framework employs a Transformer architecture to capture both global and local feature correlations across multiple imaging modalities, including MRI, CT, PET, and X-ray, for more accurate diagnostic results and faster object detection in medical imagery. The fusion process incorporates spatial and frequency-domain information to improve the clarity and detail of the output images, enhancing diagnostic accuracy. The adaptive attention mechanism within the Transformer dynamically adjusts to the relevant features of different image types, optimizing fusion in real time. This leads to an improved sensitivity (98.5%) and specificity (96.7%) in medical diagnosis. Additionally, the model significantly reduces false positives and negatives, with an F1 score of 97.2% in object detection tasks. The AT-MMIF framework is further optimized for real-time processing with an average inference time of 120 ms per image and a model size reduction of 35% compared to existing multi-modal fusion models. By leveraging the strengths of Transformer architectures and adaptive learning, the proposed framework offers a highly efficient and scalable solution for real-time medical diagnosis and object detection in various clinical settings, including radiology, oncology, and pathology.

An Algorithm for Clustering with Confidence-Based Must-Link and Cannot-Link Constraints

We study here the semisupervised k-clustering problem where information is available on whether pairs of objects are in the same or different clusters. This information is available either with certainty or with a limited level of confidence. We introduce the pair-wise confidence constraints clustering (PCCC) algorithm, which iteratively assigns objects to clusters while accounting for the information provided on the pairs of objects. Our algorithm uses integer programming for the assignment of objects, which allows us to include relationships as hard constraints that are guaranteed to be satisfied or as soft constraints that can be violated subject to a penalty. This flexibility distinguishes our algorithm from the state of the art, in which all pair-wise constraints are considered hard or all are considered soft. We developed an enhanced multistart approach and a model-size reduction technique for the integer program that contribute to the effectiveness and efficiency of the algorithm. Unlike existing algorithms, our algorithm scales to large-scale instances with up to 60,000 objects, 100 clusters, and millions of cannot-link constraints (which are the most challenging constraints to incorporate). We compare the PCCC algorithm with state-of-the-art approaches in an extensive computational study. Even though the PCCC algorithm is more general than the state-of-the-art approaches in its applicability, it outperforms the state-of-the-art approaches on instances with all hard or all soft constraints in terms of both run time and various metrics of solution quality. The code of the PCCC algorithm is publicly available on GitHub. History: Accepted by Ram Ramesh, Area Editor for Data Science and Machine Learning. Funding: The research of D. S. Hochbaum was supported by the AI Institute NSF Award [Grant 2112533]. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2023.0419 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2023.0419 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .

Lightweight Detection of Broccoli Heads in Complex Field Environments Based on LBDC-YOLO

Robotically selective broccoli harvesting requires precise lightweight detection models to efficiently detect broccoli heads. Therefore, this study introduces a lightweight and high-precision detection model named LBDC-YOLO (Lightweight Broccoli Detection in Complex Environment—You Look Only Once), based on the improved YOLOv8 (You Look Only Once, Version 8). The model incorporates the Slim-neck design paradigm based on GSConv to reduce computational complexity. Furthermore, Triplet Attention is integrated into the backbone network to capture cross-dimensional interactions between spatial and channel dimensions, enhancing the model’s feature extraction capability under multiple interfering factors. The original neck network structure is replaced with a BiFPN (Bidirectional Feature Pyramid Network), optimizing the cross-layer connection structure, and employing weighted fusion methods for better integration of multi-scale features. The model undergoes training and testing on a dataset constructed in real field conditions, featuring broccoli images under various influencing factors. Experimental results demonstrate that LBDC-YOLO achieves an average detection accuracy of 94.44% for broccoli. Compared to the original YOLOv8n, LBDC-YOLO achieves a 32.1% reduction in computational complexity, a 47.8% decrease in parameters, a 44.4% reduction in model size, and a 0.47 percentage point accuracy improvement. When compared to models such as YOLOv5n, YOLOv5s, and YOLOv7-tiny, LBDC-YOLO exhibits higher detection accuracy and lower computational complexity, presenting clear advantages for broccoli detection tasks in complex field environments. The results of this study provide an accurate and lightweight method for the detection of broccoli heads in complex field environments. This work aims to inspire further research in precision agriculture and to advance knowledge in model-assisted agricultural practices.

Construction Jobsite Image Classification Using an Edge Computing Framework

Image classification is increasingly being utilized on construction sites to automate project monitoring, driven by advancements in reality-capture technologies and artificial intelligence (AI). Deploying real-time applications remains a challenge due to the limited computing resources available on-site, particularly on remote construction sites that have limited telecommunication support or access due to high signal attenuation within a structure. To address this issue, this research proposes an efficient edge-computing-enabled image classification framework for support of real-time construction AI applications. A lightweight binary image classifier was developed using MobileNet transfer learning, followed by a quantization process to reduce model size while maintaining accuracy. A complete edge computing hardware module, including components like Raspberry Pi, Edge TPU, and battery, was assembled, and a multimodal software module (incorporating visual, textual, and audio data) was integrated into the edge computing environment to enable an intelligent image classification system. Two practical case studies involving material classification and safety detection were deployed to demonstrate the effectiveness of the proposed framework. The results demonstrated the developed prototype successfully synchronized multimodal mechanisms and achieved zero latency in differentiating materials and identifying hazardous nails without any internet connectivity. Construction managers can leverage the developed prototype to facilitate centralized management efforts without compromising accuracy or extra investment in computing resources. This research paves the way for edge “intelligence” to be enabled for future construction job sites and promote real-time human-technology interactions without the need for high-speed internet.

Lightweight rail surface defect detection algorithm based on an improved YOLOv8

Existing deep learning methods for rail surface defect detection face issues such as poor compatibility with embedded detection systems, high computational resource consumption, and slow detection speeds. To address these challenges, we propose a lightweight rail surface defect detection algorithm11Code: https://github.com/haichao67/GD-YOLOv8. based on an improved YOLOv8.Inspired by Huawei’s Gold-YOLO, our algorithm redesigns the neck architecture, integrates feature extraction, fusion, and injection modules, and replaces Complete Intersection over Union with Scylla Intersection over Union. Comparative experiments confirm its efficacy, achieving 94.9% precision and 81.9% recall on the Rail Surface Defect Dataset, a 26.3% increase in frames per second, and a model size reduction to 63%. Our algorithm not only ensures satisfactory detection results but also reduces the number of model parameters and increases the detection speed.

Improved Road Target Detection Algorithm for YOLOv7-Tiny

In complex road scenes, we propose an enhanced road target detection algorithm for YOLOv7-Tiny to address issues such as large model size, misclassification, and low localization accuracy. Our approach involves several key modifi-cations. Firstly, we replace the LeakyReLU activation function in YOLOv7-Tiny with H-swish. This replacement not only reduces the number of parameters in the model but also enhances its feature extraction capabilities. Additionally, we replace the ELAN module in the Neck with the DPCH-ELAN module and introduce the darknetblock module along with the pcconv convolution. These modifications improve the network's ability to comprehend complex patterns and semantics, thereby enabling it to capture features at different levels of input data. Moreover, we introduce the pcconv convolution building block at the output side to handle heterogeneous information in complex road scenes, thereby enhancing the network's performance in detecting road targets and abnormalities. In our experiments using a mul-ti-source dataset, the improved model exhibits a reduction in GFLOPs by 20.90% and a decrease in the number of pa-rameters by 24.49%. Furthermore, the mean average precision scores (map) at thresholds of 0.5 and 0.5~0.9 are im-proved to 77.3% and 51.8%, respectively, compared to the original YOLOv7-Tiny model. These experimental results demonstrate that our enhanced model achieves a reduction in model size while simultaneously enhancing detection accuracy, thereby meeting the requirements for real-time detection. To assess the generalizability of our approach, we conducted comparison experiments on the VOC2012 dataset. The results indicate that the improved algorithm exhibits robust generalization capabilities across different datasets.

Saliency and location aware pruning of deep visual detectors for autonomous driving

Despite the remarkable achievements of deep neural networks, their high computational complexity limits their wide use in many real-world embedded applications, such as autonomous driving perception. While current neural network pruning approaches can reduce model complexity to various extents, they often adopt local or ad hoc importance measures that are not directly related to the final task. More importantly, most of them focus on classification tasks and do not take location information into consideration during pruning. To address these issues, we present a novel channel importance measure that incorporates detection-related saliency and location awareness, specifically designed for the pruning of self-driving visual detectors. Our comprehensive experiments on the KITTI and COCO_traffic datasets demonstrate that our pruning method achieves significant reductions in model size and computational operations with little performance degradation. Notably, it outperforms other state-of-the-art methods across various pruning rates and base detectors. Our pruned YOLOX-S model with 40.2% fewer parameters even improves the original model’s mAP by 1.8% on KITTI. Moreover, we experimentally highlight the potential of our pruning approach in effectively detecting small-scale objects.

Dress Code Monitoring Method in Industrial Scene Based on Improved YOLOv8n and DeepSORT.

Deep learning-based object detection has become a powerful tool in dress code monitoring. However, even state-of-the-art detection models inevitably suffer from false alarms or missed detections, especially when handling small targets such as hats and masks. To overcome these limitations, this paper proposes a novel method for dress code monitoring using an improved YOLOv8n model, the DeepSORT tracking, and a new dress code judgment criterion. We improve the YOLOv8n model through three means: (1) a new neck structure named FPN-PAN-FPN (FPF) is introduced to enhance the model's feature fusion capability, (2) Receptive-Field Attention convolutional operation (RFAConv) is utilized to better capture the difference in information brought by different positions, and a (3) Focused Linear Attention (FLatten) mechanism is added to expand the model's receptive field. This improved YOLOv8n model increases mAP while reducing model size. Next, DeepSORT is integrated to obtain instance information across multi-frames. Finally, we adopt a new judgment criterion to conduct real-scene dress code monitoring. The experimental results show that our method effectively identifies instances of dress violations, reduces false alarms, and improves accuracy.

Investigating Hallucinations in Pruned Large Language Models for Abstractive Summarization

Abstract Despite the remarkable performance of generative large language models (LLMs) on abstractive summarization, they face two significant challenges: their considerable size and tendency to hallucinate. Hallucinations are concerning because they erode reliability and raise safety issues. Pruning is a technique that reduces model size by removing redundant weights, enabling more efficient sparse inference. Pruned models yield downstream task performance comparable to the original, making them ideal alternatives when operating on a limited budget. However, the effect that pruning has upon hallucinations in abstractive summarization with LLMs has yet to be explored. In this paper, we provide an extensive empirical study across five summarization datasets, two state-of-the-art pruning methods, and five instruction-tuned LLMs. Surprisingly, we find that hallucinations are less prevalent from pruned LLMs than the original models. Our analysis suggests that pruned models tend to depend more on the source document for summary generation. This leads to a higher lexical overlap between the generated summary and the source document, which could be a reason for the reduction in hallucination risk.1

On-Edge Deployment of Vision Transformers for Medical Diagnostics Using the Kvasir-Capsule Dataset

This paper aims to explore the possibility of utilizing vision transformers (ViTs) for on-edge medical diagnostics by experimenting with the Kvasir-Capsule image classification dataset, a large-scale image dataset of gastrointestinal diseases. Quantization techniques made available through TensorFlow Lite (TFLite), including post-training float-16 (F16) quantization and quantization-aware training (QAT), are applied to achieve reductions in model size, without compromising performance. The seven ViT models selected for this study are EfficientFormerV2S2, EfficientViT_B0, EfficientViT_M4, MobileViT_V2_050, MobileViT_V2_100, MobileViT_V2_175, and RepViT_M11. Three metrics are considered when analyzing a model: (i) F1-score, (ii) model size, and (iii) performance-to-size ratio, where performance is the F1-score and size is the model size in megabytes (MB). In terms of F1-score, we show that MobileViT_V2_175 with F16 quantization outperforms all other models with an F1-score of 0.9534. On the other hand, MobileViT_V2_050 trained using QAT was scaled down to a model size of 1.70 MB, making it the smallest model amongst the variations this paper examined. MobileViT_V2_050 also achieved the highest performance-to-size ratio of 41.25. Despite preferring smaller models for latency and memory concerns, medical diagnostics cannot afford poor-performing models. We conclude that MobileViT_V2_175 with F16 quantization is our best-performing model, with a small size of 27.47 MB, providing a benchmark for lightweight models on the Kvasir-Capsule dataset.

LiteFer: An Approach Based on MobileViT Expression Recognition.

Facial expression recognition using convolutional neural networks (CNNs) is a prevalent research area, and the network's complexity poses obstacles for deployment on devices with limited computational resources, such as mobile devices. To address these challenges, researchers have developed lightweight networks with the aim of reducing model size and minimizing parameters without compromising accuracy. The LiteFer method introduced in this study incorporates depth-separable convolution and a lightweight attention mechanism, effectively reducing network parameters. Moreover, through comprehensive comparative experiments on the RAFDB and FERPlus datasets, its superior performance over various state-of-the-art lightweight expression-recognition methods is evident.

Fabric Defect Detection Based on Improved Lightweight YOLOv8n

In response to the challenges posed by complex background textures and limited hardware resources in fabric defect detection, this study proposes a lightweight fabric defect detection algorithm based on an improved GSL-YOLOv8n model. Firstly, to reduce the parameter count and complexity of the YOLOv8n network, the GhostNet concept is used to construct the C2fGhost module, replacing the conventional convolution layers in the YOLOv8n structure with Ghost convolutions. Secondly, the SimAM parameter-free attention mechanism is embedded at the end of the backbone network to eliminate redundant background, enhance semantic information for small targets, and improve the network’s feature extraction capability. Lastly, a lightweight shared convolution detection head is designed, employing the scale layer to adjust features, ensuring the lightweight nature of the model while minimizing precision loss. Compared to the original YOLOv8n model, the improved GSL-YOLOv8n algorithm increases the mAP@0.5 by 0.60% to 98.29% and reduces model size, computational load, and parameter count by 66.7%, 58.0%, and 67.4%, respectively, meeting the application requirements for fabric defect detection in textile industry production.

Lightweight deep neural network for radio frequency interference detection and segmentation in synthetic aperture radar

Radio frequency interference (RFI) poses challenges in the analysis of synthetic aperture radar (SAR) images. Existing RFI suppression systems rely on prior knowledge of the presence of RFI. This paper proposes a lightweight neural network-based algorithm for detecting and segmenting RFI (LDNet) in the time-frequency domain. The network accurately delineates RFI pixel regions in time-frequency spectrograms. To mitigate the impact on the operational speed of the entire RFI suppression system, lightweight modules and pruning operations are introduced. Compared to threshold-based RFI detection algorithms, deep learning-based segmentation networks, and AC-UNet specifically designed for RFI detection, LDNet achieves improvements in mean intersection over union (MIoU) by 24.56%, 13.29%, and 7.54%, respectively.Furthermore, LDNet reduces model size by 99.03% and inference latency by 24.53% compared to AC-UNet.

Improved Lightweight YOLOv5 Algorithm for Detecting Grid Inspection Object

Abstract In the field of unmanned aerial vehicle-based transmission line inspection, the neural network models currently employed suffer from large model sizes and high computational requirements, resulting in a trade-off between detection speed and accuracy. We propose an improved lightweight YOLOv5 object detection algorithm based on channel pruning to address this issue. The approach incorporates the SimAM attention mechanism into the backbone network to enhance feature extraction capability without increasing computational complexity, thereby improving detection accuracy. Traditional convolutional modules in the neck network are replaced with GSconv modules to reduce convolutional computation and enhance detection speed. The Wise-IoU loss is utilized as the localization loss for the detection model, speeding up model convergence. Finally, a model compression technique is implemented to eliminate unnecessary channels within the network, resulting in a decrease in model size and a boost in detection speed. Experimental results on a constructed grid inspection dataset demonstrate that the improved model achieves a 72.1% reduction in model size, a 62.5% reduction in GFLOPs, a 5.9 FPS improvement, and a 0.3% mAP increase compared to the original YOLOv5. These findings underscore the advantages of the improved lightweight YOLOv5 object detection algorithm, offering a blend of heightened accuracy, streamlined size, and swift detection capabilities.

Field Obstacle Detection and Location Method Based on Binocular Vision

When uncrewed agricultural machinery performs autonomous operations in the field, it inevitably encounters obstacles such as persons, livestock, poles, and stones. Therefore, accurate recognition of obstacles in the field environment is an essential function. To ensure the safety and enhance the operational efficiency of autonomous farming equipment, this study proposes an improved YOLOv8-based field obstacle detection model, leveraging depth information obtained from binocular cameras for precise obstacle localization. The improved model incorporates the Large Separable Kernel Attention (LSKA) module to enhance the extraction of field obstacle features. Additionally, the use of a Poly Kernel Inception (PKI) Block reduces model size while improving obstacle detection across various scales. An auxiliary detection head is also added to improve accuracy. Combining the improved model with binocular cameras allows for the detection of obstacles and their three-dimensional coordinates. Experimental results demonstrate that the improved model achieves a mean average precision (mAP) of 91.8%, representing a 3.4% improvement over the original model, while reducing floating-point operations to 7.9 G (Giga). The improved model exhibits significant advantages compared to other algorithms. In localization accuracy tests, the maximum average error and relative error in the 2–10 m range for the distance between the camera and five types of obstacles were 0.16 m and 2.26%. These findings confirm that the designed model meets the requirements for obstacle detection and localization in field environments.

Data-driven Bayesian Gaussian mixture optimized anchor box model for accurate and efficient detection of green citrus

The demand for harvesting green citrus has increased due to the rapid growth of the green citrus industry. However, real-time detection of green citrus is challenging because of its high visual similarity to the background. To address this issue, a new model called EDGC-YOLO (Efficient Detection of Green Citrus based on the YOLO) has been developed, focusing on detecting green citrus effectively and accurately. Specifically, the proposed model uses a data-driven approach that employs the highest confidence level from the aspect ratio of unobstructed citrus fruits to infer anchor boxes of obscured citrus. This method ensures a high degree of coincidence with bounding boxes based on empirical annotation. Furthermore, analyzing the distribution range of various citrus annotation categories within the dataset and the incorporating the highest confidence level of the normal distribution regarding aspect ratios, the initial cluster centers of the BGMM (Bayesian Gaussian Mixture Model) are dynamically adjusted. This strategy enables the anchor boxes to be inferred from the best suited for detecting green citrus. These improved anchor boxes provide a more accurate and effective solution for centroid positioning of obscured green citrus, significantly enhancing the entire network model’s detection accuracy and lightweight efficiency. To further improve detection accuracy and reduce model size, this study integrated the Refined-EfficientNetV2 network as the backbone, enhanced with a Convolutional Block Attention Module (CBAM) for better feature extraction. This integration allows the model to effectively capture relevant channel information and spatial features, increasing the feature extraction capability for green citrus images. Experimental results demonstrate the model’s superior performance: parameters reduced to 4.52 million, computational demand to 7.8 GFLOPs (64.4 % and 49.4 % of the original model, respectively), and model size decreased to 9.4 MB (65.3 % of the original). Additionally, the optimized model improved Precision (P), Recall (R), and mean Average Precision (mAP) by 0.5 %, 1.6 %, and 3.0 %, respectively, compared to the original model. The proposed model achieves higher detection accuracy with a smaller model size, providing theoretical support for the harvesting decision-making of green citrus harvesting robots.

Detach-ROCKET: sequential feature selection for time series classification with random convolutional kernels

Time Series Classification (TSC) is essential in fields like medicine, environmental science, and finance, enabling tasks such as disease diagnosis, anomaly detection, and stock price analysis. While machine learning models like Recurrent Neural Networks and InceptionTime are successful in numerous applications, they can face scalability issues due to computational requirements. Recently, ROCKET has emerged as an efficient alternative, achieving state-of-the-art performance and simplifying training by utilizing a large number of randomly generated features from the time series data. However, many of these features are redundant or non-informative, increasing computational load and compromising generalization. Here we introduce Sequential Feature Detachment (SFD) to identify and prune non-essential features in ROCKET-based models, such as ROCKET, MiniRocket, and MultiRocket. SFD estimates feature importance using model coefficients and can handle large feature sets without complex hyperparameter tuning. Testing on the UCR archive shows that SFD can produce models with better test accuracy using only 10% of the original features. We named these pruned models Detach-ROCKET. We also present an end-to-end procedure for determining an optimal balance between the number of features and model accuracy. On the largest binary UCR dataset, Detach-ROCKET improves test accuracy by 0.6% while reducing features by 98.9%. By enabling a significant reduction in model size without sacrificing accuracy, our methodology improves computational efficiency and contributes to model interpretability. We believe that Detach-ROCKET will be a valuable tool for researchers and practitioners working with time series data, who can find a user-friendly implementation of the model at https://github.com/gon-uri/detach_rocket.

Improved lightweight infrared road target detection method based on YOLOv8

Infrared-based road scene object detection algorithms often face issues with excessive parameters and computational demands, making them incompatible with edge devices having constrained computational capabilities. This paper introduces an enhanced lightweight infrared-based road object detection algorithm based on YOLOv8n. Firstly, a streamlined network architecture is devised by merging YOLOv8n’s C2f module with PConv, creating a lighter module and reducing the neural network’s downsampling rate of infrared images. This strategy reduces redundant computations and memory access, preventing the loss of fine details in infrared images caused by deep convolutional neural networks. Additionally, the model’s accuracy in detecting infrared targets is significantly enhanced through the integration of the coordinate attention mechanism. Finally, replacing CIoU with Wise-IoU for bounding box regression in YOLOv8n accelerates the model’s convergence. Empirical findings indicate that in contrast to the YOLOv8n algorithm, the optimized model showcases a 34.17 % reduction in model size, a 40.35 % decrease in parameters, and a 4.8 % increase in average detection accuracy. This enhanced algorithm not only achieves a lightweight profile but also delivers superior performance on embedded edge devices.

TP-Transfiner: high-quality segmentation network for tea pest.

Detecting and controlling tea pests promptly are crucial for safeguarding tea production quality. Due to the insufficient feature extraction ability of traditional CNN-based methods, they face challenges such as inaccuracy and inefficiency of detecting pests in dense and mimicry scenarios. This study proposes an end-to-end tea pest detection and segmentation framework, TeaPest-Transfiner (TP-Transfiner), based on Mask Transfiner to address the challenge of detecting and segmenting pests in mimicry and dense scenarios. In order to improve the feature extraction inability and weak accuracy of traditional convolution modules, this study proposes three strategies. Firstly, a deformable attention block is integrated into the model, which consists of deformable convolution and self-attention using the key content only term. Secondly, the FPN architecture in the backbone network is improved with a more effective feature-aligned pyramid network (FaPN). Lastly, focal loss is employed to balance positive and negative samples during the training period, and parameters are adapted to the dataset distribution. Furthermore, to address the lack of tea pest images, a dataset called TeaPestDataset is constructed, which contains 1,752 images and 29 species of tea pests. Experimental results on the TeaPestDataset show that the proposed TP-Transfiner model achieves state-of-the-art performance compared with other models, attaining a detection precision (AP50) of 87.211% and segmentation performance of 87.381%. Notably, the model shows a significant improvement in segmentation average precision (mAP) by 9.4% and a reduction in model size by 30% compared to the state-of-the-art CNN-based model Mask R-CNN. Simultaneously, TP-Transfiner's lightweight module fusion maintains fast inference speeds and a compact model size, demonstrating practical potential for pest control in tea gardens, especially in dense and mimicry scenarios.