In the current digital era, image processing and pattern recognition play a critical role in fields such as environmental monitoring in the Internet of Things and smart city planning. However, traditional target detection algorithms face performance challenges when dealing with rotated objects, such as low recognition accuracy. To address this issue, this study investigated a deep learning-based method for detecting rotated objects. Firstly, by applying hue, saturation, value (HSV) and gamma correction to the images, the image quality was optimized to enhance object recognition capability. Secondly, this research introduced the MMRotate framework dedicated to detecting rotated objects, which, compared to traditional target detection frameworks, better meets the detection needs for rotated objects, thereby improving detection accuracy and robustness. Finally, from the perspective of the Internet of Things, this study classified and experimentally validated relevant datasets in an IoT environment, showing the performance of diferent targets on diferent datasets. Overall, this study provides new ideas and methods for addressing the shortcomings of rotated object detection in image processing and pattern recognition in the IoT environment, ofering valuable insights and guidance for the development of smart cities and environmental monitoring.
The Internet of Things (IoT) technology represents a pivotal direction in modern information
technology development, enabling various devices and objects to connect and exchange data
through embedded sensors, software, and other technologies. The widespread adoption of
this technology has profoundly transformed multiple sectors, including industry, agriculture,
healthcare, urban management, and domestic life. With the rapid development of the IoT
industry, a wide array of IoT devices has become ubiquitous in everyday life, particularly image
acquisition devices, which play a crucial role in numerous application scenarios. However,
the precision of these devices in recognizing rotated targets is often limited. Concurrently, as
the IoT scales up rapidly, the influx of massive volumes of image data into IoT environments
poses a challenge due to slow recognition speeds, which is an urgent issue that needs to be
addressed[
In recent years, as detection tasks have continuously evolved, traditional horizontal bounding boxes have become inadequate for meeting the demands of specific fields. Consequently, researchers have begun to reconsider the representation of objects. To address this challenge, methods such as increasing the degrees of freedom in regression have been adopted to achieve more flexible object detection. This novel detection approach is known as rotated object detection. Ensuring high precision while efectively conducting rotated object detection has become a current research hotspot. Numerous factors can influence the performance of deep learning-based detectors.
In today’s digital age, image processing and pattern recognition technologies have become
core components of Internet of Things (IoT) applications. With the widespread adoption of
IoT technologies, the importance of high-resolution images in urban planning, environmental
monitoring, and intelligent transportation is increasingly highlighted. However, as detection
tasks continue to evolve, traditional horizontal bounding boxes have become inadequate for
meeting the demands of specific fields. Accurately detecting and recognizing targets at various
angles and postures within images remains challenging. To address this challenge, methods
such as increasing the degrees of freedom in regression have been adopted to achieve more
lfexible object detection, known as rotated object detection. In this study, the dataset was
enhanced using Gamma correction[
The content of this paper is shown as follows. Section 2 of this paper reviews recent related work on object detection in the Internet of Things. Section 3 introduces the Gamma image processing method based on HSV and discusses the application of the MMRotate framework built upon this method. Section 4 describes how we set up the experiments and present the results. Section 5 concludes with a brief summary and mentions of future work.
The application of Internet of Things (IoT) object recognition refers to the use of IoT technology
and image processing techniques to automatically detect and identify target objects in various
scenarios. This technology demonstrates significant potential in enhancing security, optimizing
resource utilization, improving eficiency, and enhancing the quality of life across diferent fields.
Mohaimenul et al. [
Rotational object detection is used for detecting rotated, tilted, and deformed objects in images
or videos, typically described by rotated boxes or polygons, and employing deep learning models
with geometric transformations for precise detection. Cai et al. [
Current rotating target detection faces challenges such as high model complexity, high computational demands, limited labeled datasets, and dificulty in handling various angles and shapes. This paper uses an HSV-based gamma correction to enhance image quality and mAP, compares diferent rotation bounding box definitions with the MMRotate framework, and integrates HRSC and DOTA datasets to explore image processing applications in the Internet of Things.
The HSV [
Due to the nonlinear nature of human visual perception of brightness, directly displayed
images may exhibit distorted contrast in dark and bright areas. Gamma correction[
= (1) correction: the horizontal axis represents input luminance, and the vertical axis represents output luminance. The blue curve shows mapping for Gamma values < 1, while the red curve shows mapping for Gamma values > 1. With Gamma < 1, image brightness increases, enhancing contrast in low luminance areas for better detail recognition in darker regions. blue curve shows the mapping for Gamma values less than 1, enhancing overall brightness and contrast in low luminance areas. The red curve represents Gamma values greater than 1.
In experimental data preprocessing, adjusting image saturation is crucial for color performance and visual quality. Existing methods have limitations: inconsistent results across diferent image types and inadequate handling of lighting variations. This study combines a novel HSVbased Gamma correction method to enhance saturation adjustment, improving visual quality and color performance. Compared to traditional methods, this approach ofers enhanced robustness and applicability, achieving stable, accurate adjustments across scenarios.
The followings are images of the dataset processed in three ways: Figure.2 (a) shows the original image, Figure.2 (b) shows the image after gamma correction, and Figure.2 (c) shows the image after gamma correction based on HSV.
MMRotate[
The Figure.3 shows that MMRotate framework primarily consists of four components: datasets, models, core, and API. The dataset component handles data loading and preprocessing, including the datasets required for training, rotation frame data augmentation pipelines, and samplers for data loading. The model component is the core of the framework, encompassing rotation detection models and loss functions. The API component provides a user-friendly interface for model training, testing, and inference. Additionally, evaluation tools and custom hooks are integral parts of the model training core.
Ensuring high precision while efectively conducting rotated object detection has become a current research hotspot. This section will explore this topic from the perspectives of research approaches and definition methods.
OC[12] is a rotation angle representation method in a Cartesian coordinate system, typically using the coordinates of the upper-left and lower-right corners of a rectangular bounding box to represent the position and orientation of the target. OC is simple, intuitive, and easy to implement. It performs well for horizontal or vertical bounding boxes but may be less accurate for rotated bounding boxes as it struggles to describe rotation angles and aspect ratio changes. Suitable for simple horizontal or vertical bounding boxes or tasks with low rotation angle requirements.
LE135[13] is a length encoding method based on the direction of the target’s principal axis, with the angle between the principal axis and the x-axis ranging from -45 degrees to 135 degrees. LE135 can more accurately represent large-angle rotated targets and performs better for such targets compared to OC. While efective for large-angle rotations, it may not perform well for certain angle ranges or occluded targets. Suitable for tasks requiring precise description of large-angle rotated targets, enhancing detection accuracy and robustness.
LE90[14] is a simplified form of LE135, using a length encoding range of -90 degrees to 90 degrees, making it a special case of LE135. LE90 is simpler and more intuitive in calculation and representation, suitable for scenarios with lower angle precision requirements. Due to its limited range, it may be less accurate for large-angle rotated targets compared to LE135. Suitable for scenarios with low angle precision requirements, simple calculation, and representation, especially in resource-constrained situations.
The DOTA dataset [15] is a comprehensive resource for remote sensing image object detection, featuring a large collection of high-resolution images from various sensors. It includes diverse object categories, scales, and complex occlusions, making it vital for advancing object detection algorithms in high-resolution imagery. This study analyzes targets like vehicles and ships across various scales in the DOTA dataset, important for military and civilian uses, to thoroughly validate detection model accuracy, Figure.4 show the details of the dataset.
Additionally, this study also utilizes the HRSC dataset [16] as a supplementary resource. The HRSC dataset is a remote sensing image dataset used for ship detection and classification. The HRSC dataset comprises high-resolution aerial images from various angles and resolutions, along with detailed annotation information for ships associated with these images. The goal of this dataset is to provide a standard benchmark for the research and evaluation of algorithms for ship detection, classification, and recognition.
The dataset features aerial images of varying resolutions, taken under diverse temporal and weather conditions from platforms like satellites and drones. Additionally, the ships in the dataset are categorized into diferent types, including various kinds of vessels like cargo ships, fishing boats, and yachts. Each ship is annotated with detailed information, including its position, orientation, and dimensions. Every vessel is accurately marked with bounding boxes and potential directions. These annotations are crucial for training and evaluating ship detection algorithms. The dataset provides a wealth of aerial images and ship annotations, equipping researchers with ample data for studying ship detection, classification, and recognition tasks.
The HRSC dataset is a vital resource for ship detection and classification research, ofering extensive image data and annotations to aid development in this field.
To evaluate the efects of gamma correction with HSV on object recognition, a comparative experiment was conducted, analyzing color space characteristics on original images and applying gamma correction to enhance visual quality. Specifically, the HSV color space was chosen for its efectiveness in preserving color information. The results are presented as follows.
The analysis showed that images processed with Gamma correction had higher contrast and richer color saturation than the originals. Moreover, using Gamma correction in the HSV color space preserved original colors while improving visual quality. Table.1 presents the results of the relevant experiments, Figure.5 and Figure.6 visually demonstrate the regression rate andaccuracy of accuracy.
In object detection tasks, choosing an appropriate rotated bounding box definition is crucial for accurately identifying targets. Traditional rectangular bounding boxes may not efectively describe rotated or tilted objects, hence adopting more flexible rotated bounding box definitions could enhance recognition performance. This study utilized the MMRotate framework and replaced three diferent rotated bounding box definitions within the same model, including varying angle ranges, aspect ratios, and even non-rectangular shapes, to thoroughly investigate their impact on the accuracy of object recognition. bounding box definition method demonstrates certain advantages. This may be because the LE90 definition’s rotated bounding box is better suited for capturing large, horizontal or nearly horizontal targets, thereby improving the recognition performance for these types of objects. In terms of average precision, LE90 might slightly lag behind LE135, but its advantages in specific scenarios are also noteworthy.
By experimentally comparing diferent rotated bounding box definition methods, we can gain a more comprehensive understanding of their performance in object detection tasks. This provides valuable reference for selecting the most suitable rotated bounding box definition for specific scenarios.
In the same model based on the MMRotate framework, studies on diferent datasets were conducted to explore the impact of various datasets on object recognition accuracy. Specifically, multiple datasets from diferent sources and with diferent characteristics were used, covering target images in various scenarios and environments. This diverse dataset selection aims to comprehensively evaluate the model’s performance in diferent contexts, thereby better understanding its robustness and applicability. This experiment, in addition to the DOTA dataset, also included the HRSC dataset. The results are as follows.
In the paper, we used the HSV-based Gamma correction method to process images, which efectively improved visual quality and enhanced both color contrast and saturation. According to the experimental results, the mean average precision (mAP) of the original dataset is 0.74, while the mAP of the dataset processed with Gamma correction in the HSV color space is 0.82. Using the MMRotate framework, we conducted comparative experiments under three diferent rotated bounding box definitions to explore the efects of diferent conditions on rotated object recognition. Moreover, by integrating the HRSC dataset with the DOTA dataset, this study not only enriches the theoretical and practical aspects of the field but also explores the potential of these image processing techniques for object recognition and processing in the Internet of Things (IoT) environment. This provides new methods and insights for enhancing the performance of devices and services.
In the future, due to the imbalance in the number of samples across diferent categories in remote sensing image datasets, we aim to enhance the samples for categories with fewer instances to achieve a more balanced distribution. Additionally due to the limitations of the experimental environment, we were unable to conduct experiments with more models. We will strive to incorporate additional models to enrich the experimental data. (CVPR), New Orleans, LA, USA, 2022, pp. 14126–14135. doi:10.1109/CVPR52688.2022. 01375. [11] J. Wang, L. Song, Z. Li, et al., End-to-end object detection with fully convolutional network, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 15849–15858. [12] W. Li, R. Shang, Z. Ju, J. Feng, S. Xu, W. Zhang, Ellipse iou loss: Better learning for rotated bounding box regression, IEEE Geoscience and Remote Sensing Letters 21 (2024) 1–5. doi:10.1109/LGRS.2023.3345881. [13] Y. Pu, Y. Wang, Z. Xia, Y. Han, Y. Wang, W. Gan, Z. Wang, S. Song, G. Huang, Adaptive rotated convolution for rotated object detection, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 6589–6600. [14] X. Yang, X. Yang, J. Yang, Q. Ming, W. Wang, Q. Tian, J. Yan, Learning high-precision bounding box for rotated object detection via kullback-leibler divergence, Advances in Neural Information Processing Systems 34 (2021) 18381–18394. [15] G.-S. Xia, et al., Dota: A large-scale dataset for object detection in aerial images, in: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 2018, pp. 3974–3983. doi:10.1109/CVPR.2018.00418. [16] Z. Liu, L. Yuan, L. Weng, Y. Yang, A high resolution optical satellite image dataset for ship recognition and some new baselines, in: Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM), Porto, Portugal, 2017.