Single Shot Multibox Detector Research Articles

Comparative Evaluation of Convolutional Neural Network Object Detection Algorithms for Vehicle Detection.

The domain of object detection was revolutionized with the introduction of Convolutional Neural Networks (CNNs) in the field of computer vision. This article aims to explore the architectural intricacies, methodological differences, and performance characteristics of three CNN-based object detection algorithms, namely Faster Region-Based Convolutional Network (R-CNN), You Only Look Once v3 (YOLO), and Single Shot MultiBox Detector (SSD) in the specific domain application of vehicle detection. The findings of this study indicate that the SSD object detection algorithm outperforms the other approaches in terms of both performance and processing speeds. The Faster R-CNN approach detected objects in images with an average speed of 5.1 s, achieving a mean average precision of 0.76 and an average loss of 0.467. YOLO v3 detected objects with an average speed of 1.16 s, achieving a mean average precision of 0.81 with an average loss of 1.183. In contrast, SSD detected objects with an average speed of 0.5 s, exhibiting the highest mean average precision of 0.92 despite having a higher average loss of 2.625. Notably, all three object detectors achieved an accuracy exceeding 99%.

AlertTrap: A Study on Object Detection in Remote Insect Trap Monitoring System Using on the Edge Deep Learning Platform

Fruit flies are one of the most harmful insect species to fruit yields. In AlertTrap, implementation of Single-Shot Multibox Detector (SSD) architecture with different state-of-the-art backbone feature extractors such as MobileNetV1 and MobileNetV2 appears to be potential solutions for the real-time detection problem. SSD-MobileNetV1 and SSD-MobileNetV2 perform well and result in AP at 0.5 of 0.957 and 1.0, respectively. You Only Look Once (YOLO) v4-tiny outperforms the SSD family with 1.0 in AP at 0.5; however, its throughput velocity is considerably slower, which shows SSD models are better candidates for real-time implementation. We also tested the models with synthetic test sets simulating expected environmental disturbances. The YOLOv4-tiny had better tolerance to these disturbances than the SSD models. The Raspberry Pi system successfully gathered environmental data and pest counts, sending them via email over 4 G. However, running the full YOLO version in real time on Raspberry Pi is not feasible, indicating the need for a lighter object detection algorithm for future research. Among model candidates, YOLOv4-tiny generally performs best, with SSD MobileNetV2 also comparable and sometimes better, especially in scenarios with synthetic disturbances. SSD models excel in processing time, enabling real-time, high-accuracy detection. TFLITE versions of SSD models also process faster than their inference graph on TPU hardware, suggesting real-time implementation on edge devices like the Google Coral Dev Board. The results demonstrate the feasibility of real-time implementation of the fruit fly detection models on edge devices with high performance. In addition, YOLOv4-tiny is shown to be the most probable candidate because YOLOv4-tiny demonstrates a robust testing performance toward citrus fruit fly detection. Nevertheless, SSD-MobileNetV2 will be the better model, considering the inference time.

3D residual attention hierarchical fusion for real-time detection of the prostate capsule

BackgroundFor prostate electrosurgery, where real-time surveillance screens are relied upon for operations, manual identification of the prostate capsule remains the primary method. With the need for rapid and accurate detection becoming increasingly urgent, we set out to develop a deep learning approach for detecting the prostate capsule using endoscopic optical images.MethodsOur method involves utilizing the Simple, Parameter-Free Attention Module(SimAM) residual attention fusion module to enhance the extraction of texture and detail information, enabling better feature extraction capabilities. This enhanced detail information is then hierarchically transferred from lower to higher levels to aid in the extraction of semantic information. By employing a forward feature-by-feature hierarchical fusion network based on the 3D residual attention mechanism, we have proposed an improved single-shot multibox detector model.ResultsOur proposed model achieves a detection precision of 83.12% and a speed of 0.014 ms on NVIDIA RTX 2060, demonstrating its effectiveness in rapid detection. Furthermore, when compared to various existing methods including Faster Region-based Convolutional Neural Network (Faster R-CNN), Single Shot Multibox Detector (SSD), EfficientDet and others, our method Attention based Feature Fusion Single Shot Multibox Detector (AFFSSD) stands out with the highest mean Average Precision (mAP) and faster speed, ranking only below You Only Look Once version 7 (YOLOv7).ConclusionsThis network excels in extracting regional features from images while retaining the spatial structure, facilitating the rapid detection of medical images.

Contraband detection of millimeter wave image for postal security checks using a spatial transformer-feature fusion network

The image resolution and contraband object detection accuracy are the two key factors for security checks based on millimeter wave imaging techniques. In this paper, a homemade real-time millimeter imaging system for small package security inspection is used to obtain about 400 raw images of envelopes containing multi-contraband objects like guns and knives. After pre-processing, spatial transformer-feature fusion (ST-FF) adapted single-shot multi-box detector (SSD) networks are used to detect the contraband objects of postal packages. The experiments reveal that the spatial-transformed-feature fusion deep learning networks demonstrate better mean average precision (mAP) performance than traditional single networks in detecting contraband objects of different scales, orientations, and distortions, and prove the great potential for security checks based on millimeter wave imaging.

Detection of oral cancer and oral potentially malignant disorders using artificial intelligence-based image analysis.

We aimed to construct an artificial intelligence-based model for detecting oral cancer and dysplastic leukoplakia using oral cavity images captured with a single-lens reflex camera. We used 1043 images of lesions from 424 patients with oral squamous cell carcinoma (OSCC), leukoplakia, and other oral mucosal diseases. An object detection model was constructed using a Single Shot Multibox Detector to detect oral diseases and their locations using images. The model was trained using 523 images of oral cancer, and its performance was evaluated using images of oral cancer (n = 66), leukoplakia (n = 49), and other oral diseases (n = 405). For the detection of only OSCC versus OSCC and leukoplakia, the model demonstrated a sensitivity of 93.9% versus 83.7%, a negative predictive value of 98.8% versus 94.5%, and a specificity of 81.2% versus 81.2%. Our proposed model is a potential diagnostic tool for oral diseases.

Dental caries detection using faster region-based convolutional neural network with residual network

Dental caries is the highest prevalent dental disease in the world by 2022. Caries can be stopped by early detection of patients through efficient screening. Previously, there have been several methods used to detect caries such as single shot multibox detector (SSD), faster region-based convolutional neural network (Faster R-CNN) and you only look once (YOLO). This research aims to develop accurate dental caries detection using Faster R-CNN. Using a dataset collected from scraping on the internet, this research is started by creating an original dataset consisting of 81 base images which are then augmented to a total of 486 images and annotated by dental health experts from Jenderal Soedirman University. Transfer learning using pre-trained Faster R-CNN residual network (ResNet)-50 and ResNet-101 model is utilized to detect and localise dental caries. The Faster R-CNN ResNet-50 model trained using the Adam optimizer produces a mean average precision (mAP) of 0.213, and those using the momentum optimizer produce a mAP of 0.177. While the Faster R-CNN ResNet-101 model trained using the Adam optimizer produces a mAP of 0.192, and those using the momentum optimizer produce a mAP of 0.004. The model trained on the dataset showed satisfactory results in detecting dental caries, especially ResNet-50 with Adam optimizer.

TQU-HG dataset and comparative study for hand gesture recognition of RGB-based images using deep learning

Hand gesture recognition has great applications in human-computer interaction (HCI), human-robot interaction (HRI), and supporting the deaf and mute. To build a hand gesture recognition model using deep learning (DL) with high results then needs to be trained on many data and in many different conditions and contexts. In this paper, we publish the TQU-HG dataset of large RGB images with low resolution (640×480) pixels, low light conditions, and fast speed (16 fps). TQU-HG dataset includes 60,000 images collected from 20 people (10 male, 10 female) with 15 gestures of both left and right hands. A comparative study with two branches: i) based on Mediapipe TML and ii) Based on convolutional neural networks (CNNs) (you only look once (YOLO); YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLO-Nas, single shot multiBox detector (SSD) VGG16, residual network (ResNet)18, ResNext50, ResNet152, ResNext50, MobileNet V3 small, and MobileNet V3 large), the architecture and operation of CNNs models are also introduced in detail. We especially fine-tune the model and evaluate it on TQU-HG and HaGRID datasets. The quantitative results of the training and testing are presented (F1-score of YOLOv8, YOLO-Nas, MobileNet V3 small, ResNet50 is 98.99%, 98.98%, 99.27%, 99.36%, respectively on the TQU-HG dataset and is 99.21%, 99.37%, 99.36%, 86.4%, 98.3%, respectively on the HaGRID dataset). The computation time of YOLOv8 is 6.19 fps on the CPU and 18.28 fps on the GPU.

Object Detection Model: An Enhanced SSD

Abstract: Object Detection plays a significant role in today’s era. Object detection plays a crucial role in various applications such as surveillance, medical images, and autonomous vehicles. The primary goal of object detection is to accurately detect the boundaries of objects of interest and classify those objects into predefined categories. In this research paper, we enhanced an object detection technique i.e. Single shot multibox detector (SSD) which is one of the top object detection technique in both aspect accuracy and speed. SSD is an object detection model that predicts object bounding boxes and class probabilities in a single forward pass. It uses multiple feature maps at different scales to cater to various object sizes and aspect ratios.The main contribution of this research is to enhance the accuracy, recall time, precision time and mean average time (mAP). The performance of this object detection technique is improved as the number of feature maps is increases, improved backbone, and enhanced activation functions, additional layers for enhanced detection and loss function modifications. An accuracy of 0.75 indicates that the model correctly predicts the class label for approximately 75% of all objects in the dataset. The enhanced model showed these results: a precision score of 0.75 implies that 75% of the objects identified by the model as positive (detected objects) are indeed true positives. A recall of 0.73 indicates that the model successfully identifies and detects 73% of all positive instances in the dataset. A mAP of 75.25 signifies the average precision across all classes, computed at different IOU thresholds (e.g., 0.5, 0.75).

Real-Time Object Detection and Recognition with Computer Vision

Abstract: Object detection, a fundamental aspect of computer vision, is essential for identifying and localizing objects within images or video frames, leveraging advancements in deep learning, particularly convolutional neural networks (CNNs), to enhance precision and speed. Its applications span diverse domains, from autonomous vehicles and surveillance systems to augmented reality and human-computer interaction. Our project focuses on engineering a real-time object detection system, integrating deep learning and computer vision methodologies. Anchored on the robust Single Shot Multibox Detector (SSD) architecture and reinforced by the efficiency and accuracy of the MobileNetV3 backbone, our system utilizes a pre-trained SSD MobileNetV3 model and comprehensive annotations from the COCO dataset to adeptly detect and recognize a wide array of objects within live video streams or archived footage. It seamlessly processes video frames from various sources, annotating detected objects in real-time to provide instant visual feedback. Offering customizable confidence thresholds and support for multiple video sources, our project showcases the transformative potential of deep learning and computer vision, advancing realtime object detection across domains like surveillance and interactive systems. By pushing the boundaries of object detection technology, our project aims to enhance safety, efficiency, and user experiences in various applications, promising to redefine the landscape of computer vision with innovation and advancement.

SIGNSense: Auditory -Optic Impairment Communication Bridge

Language experts have recognized sign languages as natural languages with the ability to convey human emotions and ideas. Translation from written language into sign videos or extraction of spoken language sentences from sign videos is the aim of sign language translation. Sign language is the principal means of communication for the deaf and hard of hearing community, which comprises 32 million children and 328 million adults worldwide who suffer from hearing impairment. However, the inability of current systems to accurately translate and transmit sign language motions in real-time prevents effective and spontaneous communication. This research provides a revolutionary technique that enables real-time recognition of ISL gestures by integrating natural language processing with cross-modal integration. The methodology uses cutting-edge methods like the Single Shot Multibox Detector (SSD) with MobileNetV2 architecture for data collection, preprocessing, model selection, and training. In real-time inference, the trained model attains an impressive 94% accuracy rate, showcasing strong performance and encouraging outcomes for enhancing communication accessibility for people with hearing impairments

ERegPose: An explicit regression based 6D pose estimation for snake-like wrist-type surgical instruments.

Accurately estimating the 6D pose of snake-like wrist-type surgical instruments is challenging due to their complex kinematics and flexible design. We propose ERegPose, a comprehensive strategy for precise 6D pose estimation. The strategy consists of two components: ERegPoseNet, an original deep neural network model designed for explicit regression of the instrument's 6D pose, and an annotated in-house dataset of simulated surgical operations. To capture rotational features, we employ an Single Shot multibox Detector (SSD)-like detector to generate bounding boxes of the instrument tip. ERegPoseNet achieves an error of 1.056mm in 3D translation, 0.073rad in 3D rotation, and an average distance (ADD) metric of 3.974mm, indicating an overall spatial transformation error. The necessity of the SSD-like detector and L1 loss is validated through experiments. ERegPose outperforms existing approaches, providing accurate 6D pose estimation for snake-like wrist-type surgical instruments. Its practical applications in various surgical tasks hold great promise.

An object detection method for catenary component images based on improved Faster R-CNN

Catenary components are an important part of electrified railways. Especially for catenary support devices, there are various types of components with significant differences in scale. According to statistical data, there is a high risk of failure for the catenary support device components during the operation of the catenary system. Therefore, in order to ensure the safe operation of the railways, it is critical to accurately locate and recognize the components in the catenary images. In this paper, we propose an improved method based on faster region-based convolutional neural networks (Faster R-CNN) framework to realize the detection and extraction of the components on the catenary support devices. Firstly, the anchor box parameters are reset using the K-means clustering method, which greatly improves the localization precision of the predicted box. Secondly, scaled exponential linear units activation function is introduced to improve the algorithm performance. Moreover, ResNet-34, the backbone of Faster R-CNN, is optimized. We design a transition structure for multi-scale filter combination convolution to avoid missing feature information and eliminate some redundant convolution structures. This modification substantially enhances the capability of the model to recognize a wide variety of component types. Finally, we conduct some control experiments comparing with single shot multibox detector and you only look once (YOLO) series (YOLOv3, YOLOv5 and YOLOv7) models. They are faster but less accurate, especially for small objects. The results show that the proposed method has better detection performance, achieving a mean average precision of 96.50% and running at 17.79 frames per second. In addition, our model has the highest average recall of 69.27%, which is 2.66% higher than the original model.

Autonomous navigation and collision prediction of port channel based on computer vision and lidar.

This study aims to enhance the safety and efficiency of port navigation by reducing ship collision accidents, minimizing environmental risks, and optimizing waterways to increase port throughput. Initially, a three-dimensional map of the port's waterway, including data on water depth, rocks, and obstacles, is generated through laser radar scanning. Visual perception technology is adopted to process and identify the data for environmental awareness. Single Shot MultiBox Detector (SSD) is utilized to position ships and obstacles, while point cloud data create a comprehensive three-dimensional map. In order to improve the optimal navigation approach of the Rapidly-Exploring Random Tree (RRT), an artificial potential field method is employed. Additionally, the collision prediction model utilizes K-Means clustering to enhance the Faster R-CNN algorithm for predicting the paths of other ships and obstacles. The results indicate that the RRT enhanced by the artificial potential field method reduces the average path length (from 500 to 430m), average time consumption (from 30 to 22s), and maximum collision risk (from 15 to 8%). Moreover, the accuracy, recall rate, and F1 score of the K-Means + Faster R-CNN collision prediction model reach 92%, 88%, and 90%, respectively, outperforming other models. Overall, these findings underscore the substantial advantages of the proposed enhanced algorithm in autonomous navigation and collision prediction in port waterways.

Tiny machine learning empowers climbing inspection robots for real-time multiobject bolt-defect detection

Ensuring the structural integrity of steel construction is critical, necessitating effective methods for detecting bolt defects. Traditional inspection methods are reliable but require significant labor and struggle with the accessibility of complex structures. This research introduces a real-time bolt-defect detection system that integrates tiny machine learning (TinyML) with a magnetic climbing robot to enhance inspection efficiency and scope. The system employs the faster objects, more objects (FOMO) algorithm optimized for edge computing on microcontrollers. This approach enables accurate bolt identification, including normal, loose, and missing bolts. The average system accuracy is 82%, with precision and recall values from 0.57 to 0.89 and 0.67 to 0.87, respectively. The system demonstrates balanced detection, evidenced by an F1 score improvement of up to 62%. The FOMO model displays compelling performance on defect detection tasks, achieving an F1 score of approximately 75%, outperforming the MobileNetV2 Single Shot Multibox Detector with Feature Pyramid Networks Lite and You Only Look Once version 5 small. The efficiency of the FOMO model is highlighted by its low hardware requirements at less than 0.1 MB of flash memory and 893.8 KB of random access memory for the 32-bit floating-point data format and is reduced for the 8-bit integer data format, with inference times of 142 ms and 86 ms, respectively. These findings contrast with the higher resource demands and slower inference time of the compared models, indicating the suitability of the FOMO model for low-capacity microcontrollers and its feasibility for real-time applications. The performance analysis across scenarios confirms the high precision (average 0.77) and recall (average 0.76), validating the robustness of the model in diverse environmental conditions. The system offers an advancement over traditional bolt-defect detection methods in accuracy and efficiency by leveraging TinyML and a magnetic climbing robot, setting new benchmarks for real-time inspection technology. The quantitative results underscore the performance of the system, presenting a scalable, cost-effective solution for improving the safety and durability of steel structures.

Establishment and validation of an artificial intelligence-based model for real-time detection and classification of colorectal adenoma

Colorectal cancer (CRC) prevention requires early detection and removal of adenomas. We aimed to develop a computational model for real-time detection and classification of colorectal adenoma. Computationally constrained background based on real-time detection, we propose an improved adaptive lightweight ensemble model for real-time detection and classification of adenomas and other polyps. Firstly, we devised an adaptive lightweight network modification and effective training strategy to diminish the computational requirements for real-time detection. Secondly, by integrating the adaptive lightweight YOLOv4 with the single shot multibox detector network, we established the adaptive small object detection ensemble (ASODE) model, which enhances the precision of detecting target polyps without significantly increasing the model's memory footprint. We conducted simulated training using clinical colonoscopy images and videos to validate the method's performance, extracting features from 1148 polyps and employing a confidence threshold of 0.5 to filter out low-confidence sample predictions. Finally, compared to state-of-the-art models, our ASODE model demonstrated superior performance. In the test set, the sensitivity of images and videos reached 87.96% and 92.31%, respectively. Additionally, the ASODE model achieved an accuracy of 92.70% for adenoma detection with a false positive rate of 8.18%. Training results indicate the effectiveness of our method in classifying small polyps. Our model exhibits remarkable performance in real-time detection of colorectal adenomas, serving as a reliable tool for assisting endoscopists.

Automatic detection of midfacial fractures in facial bone CT images using deep learning-based object detection models

BackgroundMidfacial fractures are among the most frequent facial fractures. Surgery is recommended within 2 weeks of injury, but this time frame is often extended because the fracture is missed on diagnostic imaging in the busy emergency medicine setting. Using deep learning technology, which has progressed markedly in various fields, we attempted to develop a system for the automatic detection of midfacial fractures. The purpose of this study was to use this system to diagnose fractures accurately and rapidly, with the intention of benefiting both patients and emergency room physicians. MethodsOne hundred computed tomography images that included midfacial fractures (e.g., maxillary, zygomatic, nasal, and orbital fractures) were prepared. In each axial image, the fracture area was surrounded by a rectangular region to create the annotation data. Eighty images were randomly classified as the training dataset (3736 slices) and 20 as the validation dataset (883 slices). Training and validation were performed using Single Shot MultiBox Detector (SSD) and version 8 of You Only Look Once (YOLOv8), which are object detection algorithms. ResultsThe performance indicators for SSD and YOLOv8 were respectively: precision, 0.872 and 0.871; recall, 0.823 and 0.775; F1 score, 0.846 and 0.82; average precision, 0.899 and 0.769. ConclusionsThe use of deep learning techniques allowed the automatic detection of midfacial fractures with good accuracy and high speed. The system developed in this study is promising for automated detection of midfacial fractures and may provide a quick and accurate solution for emergency medical care and other settings.

Dual attention-enhanced SSD: A novel deep learning model for object detection

Object detection is a fundamental task in computer vision with significant implications across various applications, including autonomous driving, surveillance, and image understanding. The accurate and efficient detection of objects within images is crucial for enabling machines to interpret visual information and make informed decisions. In this paper, we present an enhanced version of the Single Shot MultiBox Detector (SSD) for object detection, leveraging the concept of dual attention mechanisms. Our proposed approach, named SSD-Dual Attention, integrates dual attention layers into the SSD framework. These dual attention layers are strategically positioned between feature maps and prediction convolutions, enhancing the model's ability to capture informative feature representations across a wide range of object scales and backgrounds. Experimental results on the PASCAL VOC 2007 and 2012 datasets validate the effectiveness of our approach. Notably, SSD-Dual Attention achieves an impressive mean Average Precision (mAP) of 78.1%, surpassing the performance of SSD models enhanced with attention mechanisms such as SSD-ECA, SSD-CBAM, SSD-Non-local attention and SSD-SE attention, as well as the original SSD. These results underscore the enhanced accuracy and precision of our object detection system, marking a substantial advancement in the field of computer vision. Code is available at https://github.com/AlexHunterLeo/Dual-attention-Enhanced-SSD-A-Novel-Deep-Learning-Model-for-Object-Detection

Image Recognition System for Mechanical Parts Based on SSD and YOLOv5

With the development of industrial automation, the importance of mechanical parts in the production process is becoming increasingly important. In recent years, deep learning technology has made breakthrough progress in fields such as image recognition and object detection. Deep learning models can learn a large amount of data through multiple epochs, automatically extract features from it, and achieve image recognition and object detection. This provides potent technical assistance for the design of mechanical component recognition systems. In industrial production, there are a wide variety of mechanical parts with complex shapes, and the differences between different models are often very small, making it difficult to identify them manually. Therefore, it is necessary to develop a system that can automatically identify mechanical parts to improve production efficiency and quality. This article takes mechanical parts as the research object and studies the recognition technology of mechanical parts based on deep learning and image recognition. The main content of the article includes the following aspects. Research was conducted on image quality enhancement methods for images with different characteristics, and different image preprocessing methods were applied to different images. YOLOv5 and SSD (Single Shot MultiBox Detector) algorithms are used to perform deep learning and recognition detection on the processed images, respectively. By comprehensively comparing the differences between the two algorithms, YOLOv5 was selected as the focus of this design, and based on this, deep training was conducted, and the code was imported into the embedded system.

Object Detection Using CNN

Object detection system using Convolutional Neural Network(CNN) that can accurately identify and classify objects in videos. The purpose of object detection using CNN to enhance technology such as security cameras, smart devices by enabling them to identify and understand objects in videos. Object detection using CNN is a fascinating filed in computer vision. Detection can be difficult since there are all kinds of variations in orientation, lighting, background that can result in completely different videos of the very same object. Now with the advance of deep learning and neural network, we can finally tackle such problems without coming up with various heuristics real-time. The project “Object detection using CNN for video streaming” detects objects efficiently based on CNN algorithm and apply the algorithm on image or video data. In this project, we develop a technique to identify an object considering the deep learning pre-trained model MobileNet for Single Shot Multi-Box Detector (SSD). This algorithm is used for real-time detection and for webcam, which detects the objects in a video stream. Therefore, we use an object detection module that can detect what is in the video stream. In order to implement the module, we combine the MobileNet and the SSD framework for a fast and efficient deep learning-based method of object detection. The main purpose of our research is to elaborate the accuracy of an object detection method SSD and the importance of pre-trained deep learning model MobileNet. The experimental results show that the Average Precision (AP) of the algorithm to detect different classes as car, person and chair is 99.76%, 97.76% and 71.07%, respectively. The main objective of our project is to make clear the object detecting accuracy. The existing methods are Region Based Convolutional Neural Network(R-CNN) and You Only Look Once(YOLO).R-CNN could not pushed real time speed though its system is updated and new versions of it are deployed and YOLO network is popular but YOLO is to struggle to detect objects grouped close together, especially smaller ones. To avoid the drawbacks of these methods we proposed this model which included single shot multi-box detector (SSD), this algorithm is used for real time detection and Mobile-Net architecture.

Floating object detection using double-labelled domain generalization

Detection of floating objects in complicated aquatic environments has a wide range of applications, but it confronts significant hurdles due to the imbalance in detection accuracy and efficiency, and low domain generalization performance. To address these issues, this study proposes a novel floating object detection method based on double-labelled domain generalization. First, the Single Shot Multibox Detector (SSD) is improved by replacing the backbone network with a lightweight feature extraction network, and dynamic feature pyramid network is introduced to balance accuracy and efficiency. Then, this study initializes the improved SSD network based on the double-labelled data of the source domain, and filters the feature extractor bias, classification bias and location bias using pseudo-labelling and feature projection based on the un-labelled source domain data to minimize the bias to improve the domain generalization performance. The proposed method is trained on a self-constructed floating object dataset and is compared with state-of-the-art methods based on multiple scenarios. The results show that the proposed method achieves better performance in double-labelled domain generalization and conventional domain generalization tasks compared to other methods, achieving 70.33%, 22.38 f/s and 85.29% and 17.81 f/s in accuracy and speed respectively, satisfying the need for multi-scale floating object detection in complex environments and also alleviating the data labelling problem. This work effectively solves the problem of slow detection of floating objects due to the complex model structure and low generalization ability, and provides support for the rapid detection of floating objects in complex scenarios and the promotion of technology applications.