Deep Learning-based Detection Research Articles

Contrasting YOLOv7, SSD, and DETR on Insulator Identification under Small-sample Learning

Background:: Daily inspections of insulators are necessary because they are indispensable components for power transmission lines. Using deep learning to monitor insulators is a newly developed method. However, most deep learning-based detection methods rely on a large training sample set, which consumes computing resources and increases the workload of sample labeling. The selection of learning models to monitor insulators becomes problematic. Objective:: Through comparative analysis, a model suitable for small-sample insulator learning is found to provide a reference for the research and application of insulator detection. Methods:: This paper compares some of the latest deep learning models, YOLOv7, SSD, and DETR, for insulator detection based on small-sample learning. The small sample here means that the number of samples and their proportion to the total sample are relatively small. Two public insulator image sets, InsulatorDataSet with 600 insulator images and Transmission-line-pictures (TLP) with 1230 insulator images in the natural background are selected to test the performance of these models. Results:: Tests on two public insulator image sets, InsulatorDataSet and TLP, show that the recognition rates of YOLOv7, DETR, and SSD are arranged from high to low. The DETR and the YOLOv7 have stable performance, while the SSD lacks stable performance in terms of the learning time and recognition rate. Conclusion:: The in-domain and cross-domain scenario tests show that YOLOv7 is more suitable for insulator detection under small-sample conditions among the three models.

Deep learning-based detection of irreversible pulpitis in primary molars.

Changes in healthy and inflamed pulp on periapical radiographs are traditionally so subtle that they may be imperceptible to human experts, limiting its potential use as an adjunct clinical diagnostic feature. This study aimed to investigate the feasibility of an image-analysis technique based on the convolutional neural network (CNN) to detect irreversible pulpitis in primary molars on periapical radiographs (PRs). This retrospective study was performed in two health centres. Patients who received indirect pulp therapy at Peking University Hospital for Stomatology were retrospectively identified and randomly divided into training and validation sets (8:2). Using PRs as input to an EfficientNet CNN, the model was trained to categorise cases into either the success or failure group and externally tested on patients who presented to our affiliate institution. Model performance was evaluated using sensitivity, specificity, accuracy and F1 score. A total of 348 PRs with deep caries were enrolled from the two centres. The deep learning model achieved the highest accuracy of 0.90 (95% confidence interval: 0.79-0.96) in the internal validation set, with an overall accuracy of 0.85 in the external test set. The mean greyscale value was higher in the failure group than in the success group (p = .013). The deep learning-based model could detect irreversible pulpitis in primary molars with deep caries on PRs. Moreover, this study provides a convenient and complementary method for assessing pulp status.

Fake User Detection Based on Multi-Model Joint Representation

The existing deep learning-based detection of fake information focuses on the transient detection of news itself. Compared to user category profile mining and detection, transient detection is prone to higher misjudgment rates due to the limitations of insufficient temporal information, posing new challenges to social public opinion monitoring tasks such as fake user detection. This paper proposes a multimodal aggregation portrait model (MAPM) based on multi-model joint representation for social media platforms. It constructs a deep learning-based multimodal fake user detection framework by analyzing user behavior datasets within a time retrospective window. It integrates a pre-trained Domain Large Model to represent user behavior data across multiple modalities, thereby constructing a high-generalization implicit behavior feature spectrum for users. In response to the tendency of existing fake user behavior mining to neglect time-series features, this study introduces an improved network called Sequence Interval Detection Net (SIDN) based on Sequence to Sequence (seq2seq) to characterize time interval sequence behaviors, achieving strong expressive capabilities for detecting fake behaviors within the time window. Ultimately, the amalgamation of latent behavioral features and explicit characteristics serves as the input for spectral clustering in detecting fraudulent users. The experimental results on Weibo real dataset demonstrate that the proposed model outperforms the detection utilizing explicit user features, with an improvement of 27.0% in detection accuracy.

Range-Spread Target Detection Networks Using HRRPs

Range-spread target (RST) detection is an important issue for high-resolution radar (HRR). Traditional detectors relying on manually designed detection statistics have their performance limitations. Therefore, in this work, two deep learning-based detectors are proposed for RST detection using HRRPs, i.e., an NLS detector and DFCW detector. The NLS detector leverages domain knowledge from the traditional detector, treating the input HRRP as a low-level feature vector for target detection. An interpretable NLS module is designed to perform noise reduction for the input HRRP. The DFCW detector takes advantage of the extracted high-level feature map of the input HRRP to improve detection performance. It incorporates a feature cross-weighting module for element-wise feature weighting within the feature map, considering the channel and spatial information jointly. Additionally, a nonlinear accumulation module is proposed to replace the conventional noncoherent accumulation operation in the double-HRRP detection scenario. Considering the influence of the target spread characteristic on detector performance, signal sparseness is introduced as a measure and used to assist in generating two datasets, i.e., a simulated dataset and measured dataset incorporating real target echoes. Experiments based on the two datasets are conducted to confirm the contribution of the designed modules to detector performance. The effectiveness of the two proposed detectors is verified through performance comparison with traditional and deep learning-based detectors.

Proactive hybrid learning framework for real-time multi-vehicle detection in unregulated traffic environments

Reliable multi-vehicle detection in unregulated traffic environments is a crucial computer vision task in the development of Intelligent Transportation Systems (ITS). Despite the promising potential of Deep Learning (DL) methods for vehicle detection, the presence of uncertainties such as varying vehicle shapes and sizes, intricate background clutter, and unpredictable vehicle flow contribute to the chaotic unregulated traffic. For real-time vehicle detection in an unregulated traffic environment this paper proposes the Hybrid Learning Multi-Vehicle Detection framework (HL-MVD) based on the Convolutional Multi-head Attention Transformer Detector (CMATDet). The primary objective of this study is to generate a dataset containing highly informative video frames using a Pool-based Active Learning Strategy (PALS). Additionally, transfer learning will be implemented to train CMATDet to achieve improved accuracy and reduced detection latency. The proposed approach restructures the baseline YOLOv5x and incorporates a Multi-head Attention transformer encoder to effectively extract global features and a Scale Specific Bidirectional Feature Pyramid Network (SS-BiFPN) to facilitate multi-scale feature representation. Simplified Optimal Transport Algorithm with top-q approximation technique (Sim-OTA) is utilized for label assignment approach. Heatmap analysis demonstrated the suitability of the newly generated dataset “AU-INV-P-PALS” for detecting specific Indian native vehicles. The performance of the proposed framework is evaluated on our custom-developed AU-INV-P-PALS vehicle dataset and the IITM-HeTra Dataset 1. In comparison with contemporary detection models, the proposed HL-MVD framework resulted with higher mAP scores (91.1% on mAP@0.5 and 78.3% on mAP@0.5:0.95) for AU-INV-P-PALS. The proposed model demonstrated lower inference latency (8.1 ms), higher precision score (82.7% for IoU = 0.5), and higher recall score (90.8% for IoU = 0.5) than recent deep learning-based detection models in the literature. The top-q approximation technique in the detection head results in a reduced false-positive rate compared to conventional models. Finally, the performance of the proposed framework is tested on CCTV traffic footage captured on city roads in Chennai, India.

Deep Learning-Based Detection of Human Blastocyst Compartments with Fractal Dimension Estimation

In vitro fertilization (IVF) is an efficacious form of aided reproduction to deal with infertility. Human embryos are taken from the body, and these are kept in a supervised laboratory atmosphere during the IVF technique until they exhibit blastocyst properties. A human expert manually analyzes the morphometric properties of the blastocyst and its compartments to predict viability through manual microscopic evaluation. A few deep learning-based approaches deal with this task via semantic segmentation, but they are inaccurate and use expensive architecture. To automatically detect the human blastocyst compartments, we propose a parallel stream fusion network (PSF-Net) that performs the semantic segmentation of embryo microscopic images with inexpensive shallow architecture. The PSF-Net has a shallow architecture that combines the benefits of feature aggregation through depth-wise concatenation and element-wise summation, which helps the network to provide accurate detection using 0.7 million trainable parameters only. In addition, we compute fractal dimension estimation for all compartments of the blastocyst, providing medical experts with significant information regarding the distributional characteristics of blastocyst compartments. An open dataset of microscopic images of the human embryo is used to evaluate the proposed approach. The proposed method also demonstrates promising segmentation performance for all compartments of the blastocyst compared with state-of-the-art methods, achieving a mean Jaccard index (MJI) of 87.69%. The effectiveness of PSF-Net architecture is also confirmed with the ablation studies.

Fully automated planning for anatomical fetal brain MRI on 0.55T.

Widening the availability of fetal MRI with fully automatic real-time planning of radiological brain planes on 0.55T MRI. Deep learning-based detection of key brain landmarks on a whole-uterus echo planar imaging scan enables the subsequent fully automatic planning of the radiological single-shot Turbo Spin Echo acquisitions. The landmark detection pipeline was trained on over 120 datasets from varying field strength, echo times, and resolutions and quantitatively evaluated. The entire automatic planning solution was tested prospectively in nine fetal subjects between 20 and 37 weeks. A comprehensive evaluation of all steps, the distance between manual and automatic landmarks, the planning quality, and the resulting image quality was conducted. Prospective automatic planning was performed in real-time without latency in all subjects. The landmark detection accuracy was 4.2 2.6 mm for the fetal eyes and 6.5 3.2 for the cerebellum, planning quality was 2.4/3 (compared to 2.6/3 for manual planning) and diagnostic image quality was 2.2 compared to 2.1 for manual planning. Real-time automatic planning of all three key fetal brain planes was successfully achieved and will pave the way toward simplifying the acquisition of fetal MRI thereby widening the availability of this modality in nonspecialist centers.

Deep learning-based detection of coronary artery disease from electrocardiogram (ECG) images

Deep learning-based detection of coronary artery disease from electrocardiogram (ECG) images

Deep learning-based detection of indicator species for monitoring biodiversity in semi-natural grasslands

Deep learning (DL) has huge potential to provide valuable insights into biodiversity changes in species-rich agricultural ecosystems such as semi-natural grasslands, helping to prioritize and plan conservation efforts. However, DL has been underexplored in grassland conservation efforts, hindered by data scarcity, intricate ecosystem interactions, and limited economic incentives. Here, we developed a DL-based object-detection model to identify indicator species, a group of vascular plant species that serve as surrogates for biodiversity assessment in high nature value (HNV) grasslands. We selected indicator species Armeria maritima, Campanula patula, Cirsium oleraceum, and Daucus carota. To overcome the hurdle of limited data, we grew indicator plants under controlled greenhouse conditions, generating a sufficient dataset for DL model training. The model was initially trained on this greenhouse dataset. Then, smaller datasets derived from an experimental grassland plot and natural grasslands were added to the training to facilitate the transition from greenhouse to field conditions. Our optimized model achieved remarkable average precision (AP) on test datasets, with 98.6 AP50 on greenhouse data, 98.2 AP50 on experimental grassland data, and 96.5 AP50 on semi-natural grassland data. Our findings highlight the innovative application of greenhouse-grown specimens for the in-situ identification of plants, bolstering biodiversity monitoring in grassland ecosystems. Furthermore, the study illuminates the promising role of DL techniques in conservation programs, particularly as a monitoring tool to support result-based agri-environment schemes.

CSDFormer: A cloud and shadow detection method for landsat images based on transformer

Cloud and shadow (CS) detection is crucial prerequisite for application of remote sensing images. Current deep learning-based detection algorithms mainly employ Convolutional Neural Networks (CNNs). However, the local receptive field in CNNs cannot effectively capture global contextual information, which hinders accurate characterization of the dependency between clouds and shadows. In vision Transformers, self-attention mechanisms can effectively capture the long-distance dependencies between different regions in an image. Inspired by this, this paper proposed a new CS Detection algorithm based on a Transformer, called CSDFormer. Specifically, we exclusively employed a hierarchical Transformer structure in the encoder stage to extract features of CS. Each Transformer layer contains several multi-head self-attention mechanisms for calculating pixel-wise long-distance connectivity. The designed structure enables the Transformer to better extract global context information, which helps to strengthen the comprehension of the semantic relationships between clouds and shadows. Benefiting from the global feature extraction capability of the encoder stage, we employed several simple multilayer perceptron layers for multi-scale feature map fusion and pixel classification in the decoder stage. The proposed CSDFormer was validated using 898 Landsat 8 Biome images with 512 × 512 pixels, producing an overall accuracy of 95.28 % and a mean intersection over union of 84.08 %, outperforming three state-of-the-art CNN-based algorithms. CSDFormer is consistently more accurate in detection of both clouds and shadows. Owing to the parallel computing capability of the self-attention mechanism, CSDFormer is computationally more efficient than the three CNN-based benchmark methods. For the input spectral bands, the performance of CSDFormer produced can be further enhanced with additional thermal infrared bands.

VulMPFF: A Vulnerability Detection Method for Fusing Code Features in Multiple Perspectives

Source code vulnerabilities are one of the significant threats to software security. Existing deep learning-based detection methods have proven their effectiveness. However, most of them extract code information on a single intermediate representation of code (IRC), which often fails to extract multiple information hidden in the code fully, significantly limiting their performance. To address this problem, we propose VulMPFF, a vulnerability detection method that fuses code features under multiple perspectives. It extracts IRC from three perspectives: code sequence, lexical and syntactic relations, and graph structure to capture the vulnerability information in the code, which effectively realizes the complementary information of multiple IRCs and improves vulnerability detection performance. Specifically, VulMPFF extracts serialized abstract syntax tree as IRC from code sequence, lexical and syntactic relation perspective, and code property graph as IRC from graph structure perspective, and uses Bi-LSTM model with attention mechanism and graph neural network with attention mechanism to learn the code features from multiple perspectives and fuse them to detect the vulnerabilities in the code, respectively. We design a dual-attention mechanism to highlight critical code information for vulnerability triggering and better accomplish the vulnerability detection task. We evaluate our approach on three datasets. Experiments show that VulMPFF outperforms existing state-of-the-art vulnerability detection methods (i.e., Rats, FlawFinder, VulDeePecker, SySeVR, Devign, and Reveal) in Acc and F1 score, with improvements ranging from 14.71% to 145.78% and 152.08% to 344.77%, respectively. Meanwhile, experiments in the open-source project demonstrate that VulMPFF has the potential to detect vulnerabilities in real-world environments.

Object-Enhanced YOLO Networks for Synthetic Aperture Radar Ship Detection

Synthetic aperture radar (SAR) enables precise object localization and imaging, which has propelled the rapid development of algorithms for maritime ship identification and detection. However, most current deep learning-based algorithms tend to increase network depth to improve detection accuracy, which may result in the loss of effective features of the target. In response to this challenge, this paper innovatively proposes an object-enhanced network, OE-YOLO, designed specifically for SAR ship detection. Firstly, we input the original image into an improved CFAR detector, which enhances the network’s ability to localize and perform object extraction by providing more information through an additional channel. Additionally, the Coordinate Attention mechanism (CA) is introduced into the backbone of YOLOv7-tiny to improve the model’s ability to capture spatial and positional information in the image, thereby alleviating the problem of losing the position of small objects. Furthermore, to enhance the model’s detection capability for multi-scale objects, we optimize the neck part of the original model to integrate the Asymptotic Feature Fusion (AFF) network. Finally, the proposed network model is thoroughly tested and evaluated using publicly available SAR image datasets, including the SAR-Ship-Dataset and HRSID dataset. In comparison to the baseline method YOLOv7-tiny, OE-YOLO exhibits superior performance with a lower parameter count. When compared with other commonly used deep learning-based detection methods, OE-YOLO demonstrates optimal performance and more accurate detection results.

TAG-Net: Target Attitude Angle-Guided Network for Ship Detection and Classification in SAR Images

Synthetic aperture radar (SAR) ship detection and classification has gained unprecedented attention due to its important role in maritime transportation. Many deep learning-based detectors and classifiers have been successfully applied and achieved great progress. However, ships in SAR images present discrete and multi-centric features, and their scattering characteristics and edge information are sensitive to variations in target attitude angles (TAAs). These factors pose challenges for existing methods to obtain satisfactory results. To address these challenges, a novel target attitude angle-guided network (TAG-Net) is proposed in this article. The core idea of TAG-Net is to leverage TAA information as guidance and use an adaptive feature-level fusion strategy to dynamically learn more representative features that can handle the target imaging diversity caused by TAA. This is achieved through a TAA-aware feature modulation (TAFM) module. It uses the TAA information and foreground information as prior knowledge and establishes the relationship between the ship scattering characteristics and TAA information. This enables a reduction in the intra-class variability and highlights ship targets. Additionally, considering the different requirements of the detection and classification tasks for the scattering information, we propose a layer-wise attention-based task decoupling detection head (LATD). Unlike general deep learning methods that use shared features for both detection and classification tasks, LATD extracts multi-level features and uses layer attention to achieve feature decoupling and select the most suitable features for each task. Finally, we introduce a novel salient-enhanced feature balance module (SFB) to provide richer semantic information and capture the global context to highlight ships in complex scenes, effectively reducing the impact of background noise. A large-scale ship detection dataset (LSSDD+) is used to verify the effectiveness of TAG-Net, and our method achieves state-of-the-art performance.

Decoding scheme based on CNN for differential free space optical communication system

A high-speed and robust decoding algorithm based on Convolutional Neural Networks (CNN) is proposed for differential free space optical communication system. The system adopts bipolar complementary pulse width modulation (BCPWM) that can enhance the amplitude of the signal and avoid interference from common mode signals in a single channel system. In addition, BCPWM has amplitude and envelope characteristics. The deep learning-based detector proposed by the signal does not require channel state information. It directly feeds the received signal into a deep neural network, and CNN can effectively extract amplitude and envelope features from the signal. The experimental comparison was conducted with bipolar non-return to zero (BNRZ) signals with only amplitude features to verify the demodulation ability of CNN for signals with different features. With a bit error rate of 10−5, the required signal-to-noise ratio for BNRZ is 13 dB, and the required signal-to-noise ratio for BCPWM is 8 dB. The demodulation performance of CNN, maximum likelihood (ML), and Volterra equalizer detector were compared. For BCPWM signals, compared to ML and Volterra equalizer detectors, the proposed CNN-based detector reduces the required signal-to-noise ratio by 2 dB and 8 dB at a bit error rate of 10−5, respectively.

All-atom RNA structure determination from cryo-EM maps.

Many methods exist for determining protein structures from cryogenic electron microscopy maps, but this remains challenging for RNA structures. Here we developed EMRNA, a method for accurate, automated determination of full-length all-atom RNA structures from cryogenic electron microscopy maps. EMRNA integrates deep learning-based detection of nucleotides, three-dimensional backbone tracing and scoring with consideration of sequence and secondary structure information, and full-atom construction of the RNA structure. We validated EMRNA on 140 diverse RNA maps ranging from 37 to 423 nt at 2.0-6.0 Å resolutions, and compared EMRNA with auto-DRRAFTER, phenix.map_to_model and CryoREAD on a set of 71 cases. EMRNA achieves a median accuracy of 2.36 Å root mean square deviation and 0.86 TM-score for full-length RNA structures, compared with 6.66 Å and 0.58 for auto-DRRAFTER. EMRNA also obtains a high residue coverage and sequence match of 93.30% and 95.30% in the built models, compared with 58.20% and 42.20% for phenix.map_to_model and 56.45% and 52.3% for CryoREAD. EMRNA is fast and can build an RNA structure of 100 nt within 3 min.

Study on deep learning-based detection of viable cell count in dialysis fluid images

Background and aimsControl of dialysate fluid quality is critical to secure the safety of dialysis treatment. The number of colonies is manually counted when determining viable cell count in dialysis fluid, but errors and subjective interpretation on the part of the measurer can be problematic. This prompted us to examine the potential for using deep learning to detect viable cells and count their numbers.MethodsIn this study we prepared 5360 images for detecting viable cell count and classified them into four categories using the VGG-16 model. These images were resized to 224 × 224 px; 90% of them were used for learning, and 10% were used for validation. In an alternative approach, we also created 110 annotated images from images to detect viable cell count in dialysis fluid and subjected them to learning using the YOLOv5 model.ResultsVGG-model had a detection accuracy using the test data was 43%. YOLOv5 model had a mAP (Mean Average Precision) was 0.842. The detection accuracy using the test data was 90%.ConclusionsThe method using the VGG-16 model had problems with overfitting, suggesting that the model was not sufficiently expressive. The detection of viable cells using the YOLOv5 model showed high accuracy.

Deep Learning-Based Detection of Defects in Wafer Buffer Zone During Semiconductor Packaging Process

AbstractEffectively detecting defects in wafer buffer zones is crucial in semiconductor manufacturing processes. If defects are not detected in the middle of the semiconductor manufacturing process, the semiconductor eventually becomes a defective product after all processes are completed. To solve this problem, rule-based vision algorithms have been used to identify defects in the wafer buffer zone. After photographing the wafer buffer zone using a high-speed camera, defects are detected using the pixel values of the images. However, because of the resin bleed in the wafer, which is an epoxy compound, it is difficult to detect defects. Therefore, we introduced a deep learning method for semiconductor inspection and created a high-performance semiconductor inspection algorithm. The defects in the wafer buffer zone should be detected accurately and quickly. Furthermore, the approximate size of the defects must be extracted. We modified the Xception model to fit the wafer data characteristics considering both accuracy and speed. We proposed to extract the size of the defect using class activation mapping (CAM). We obtained an accuracy of 96.9% from the actual wafer dataset through this framework, and then managed to extract the size of the defect through CAM.

Deep Learning-Based Detection of Learning Disorders on a Large Scale Dataset of Eye Movement Records

Early detection of dyslexia and learning disorders is vital for avoiding a learning disability, as well as supporting dyslexic students by tailoring academic programs to their needs. Several studies have investigated using supervised algorithms to screen dyslexia vs. control subjects; however, the data size and the conditions of data acquisition were their most significant limitation. In the current study, we leverage a large dataset, containing 4243 time series of eye movement records from children across Europe. These datasets were derived from various tests such as saccade, vergence, and reading tasks. Furthermore, our methods were evaluated with realistic test data, including real-life biases such as noise, eye tracking misalignment, and similar pathologies among non-scholar difficulty classes. In addition, we present a novel convolutional neural network architecture, adapted to our time series classification problem, that is intended to generalize on a small annotated dataset and to handle a high-resolution signal (1024 point). Our architecture achieved a precision of 80.20% and a recall of 75.1%, when trained on the vergence dataset, and a precision of 77.2% and a recall of 77.5% when trained on the saccade dataset. Finally, we performed a comparison using our ML approach, a second architecture developed for a similar problem, and two other methods that we investigated that use deep learning algorithms to predict dyslexia.

Towards Fully Autonomous UAV: Damaged Building-Opening Detection for Outdoor-Indoor Transition in Urban Search and Rescue

Autonomous unmanned aerial vehicle (UAV) technology is a promising technology for minimizing human involvement in dangerous activities like urban search and rescue missions (USAR), both in indoor and outdoor. Automated navigation from outdoor to indoor environments is not trivial, as it encompasses the ability of a UAV to automatically map and locate the openings in a damaged building. This study focuses on developing a deep learning model for the detection of damaged building openings in real time. A novel damaged building-opening dataset containing images and mask annotations, as well as a comparison between single and multi-task learning-based detectors are given. The deep learning-based detector used in this study is based on YOLOv5. First, this study compared the different versions of YOLOv5 (i.e., small, medium, and large) capacity to perform damaged building-opening detections. Second, a multitask learning YOLOv5 was trained on the same dataset and compared with the single-task detector. The multitask learning (MTL) was developed based on the YOLOv5 object detection architecture, adding a segmentation branch jointly with the detection head. This study found that the MTL-based YOLOv5 can improve detection performance by combining detection and segmentation losses. The YOLOv5s-MTL trained on the damaged building-opening dataset obtained 0.648 mAP, an increase of 0.167 from the single-task-based network, while its inference speed was 73 frames per second on the tested platform.

Intrarow Uncut Weed Detection Using You-Only-Look-Once Instance Segmentation for Orchard Plantations.

Mechanical weed management is a drudging task that requires manpower and has risks when conducted within rows of orchards. However, intrarow weeding must still be conducted by manual labor due to the restricted movements of riding mowers within the rows of orchards due to their confined structures with nets and poles. However, autonomous robotic weeders still face challenges identifying uncut weeds due to the obstruction of Global Navigation Satellite System (GNSS) signals caused by poles and tree canopies. A properly designed intelligent vision system would have the potential to achieve the desired outcome by utilizing an autonomous weeder to perform operations in uncut sections. Therefore, the objective of this study is to develop a vision module using a custom-trained dataset on YOLO instance segmentation algorithms to support autonomous robotic weeders in recognizing uncut weeds and obstacles (i.e., fruit tree trunks, fixed poles) within rows. The training dataset was acquired from a pear orchard located at the Tsukuba Plant Innovation Research Center (T-PIRC) at the University of Tsukuba, Japan. In total, 5000 images were preprocessed and labeled for training and testing using YOLO models. Four versions of edge-device-dedicated YOLO instance segmentation were utilized in this research-YOLOv5n-seg, YOLOv5s-seg, YOLOv8n-seg, and YOLOv8s-seg-for real-time application with an autonomous weeder. A comparison study was conducted to evaluate all YOLO models in terms of detection accuracy, model complexity, and inference speed. The smaller YOLOv5-based and YOLOv8-based models were found to be more efficient than the larger models, and YOLOv8n-seg was selected as the vision module for the autonomous weeder. In the evaluation process, YOLOv8n-seg had better segmentation accuracy than YOLOv5n-seg, while the latter had the fastest inference time. The performance of YOLOv8n-seg was also acceptable when it was deployed on a resource-constrained device that is appropriate for robotic weeders. The results indicated that the proposed deep learning-based detection accuracy and inference speed can be used for object recognition via edge devices for robotic operation during intrarow weeding operations in orchards.