Deep Learning Object Detection Model Research Articles

Comprehensive study on the development of an automatic helmet violator detection system (AHVDS) using advanced machine learning techniques

The number of motorcycle accidents has increased rapidly due to different causes. Helmet is a one of the major safety equipment that helps to reduce the fatal accidents even though many of the motorcyclists are not used that. Currently many researchers have focused on traffic violator detection but they are yet to identify the violated person details as it requires high resolution cameras and expensive sensors for number plate extraction. This research article is focused on comprehensive study on the development of an automated helmet violation detection with number plate extraction in real time. This study divides the dataset into rider with helmet, rider without helmet, passenger with helmet and passenger without helmet classes, and four training model was created to detect helmet in rider and passengers. The proposed model was constructed in three stages to attain high accuracy with a vast amount of training dataset. In the first stage, the helmet was detected in the input image using the deep learning object-detection model .In the second stage, if the helmet identification confirmed in the input image then the entire motorcyclist image is segmented based on object segmentation model. In the third stage, number plate training model is deployed into the same object detection and object segmentation model to identify and segment the license plates in the motorcyclist image. This study primarily incorporates the Mask Region based Convolution Neural Networks (Mask R-CNN) for object detection and segmentation Further, it incorporates the three different deep neural networks like Dense block Network (DenseNet-121), Residual Network (ResNet-101) and Inception Residual Network architecture V2layers (Inception Res Net V2) to create accurate helmet violator detections with more extensive information. The experimental results indicates that this comprehensive helmet violator detection study improves the helmet detection accuracy by 6.2 %.In addition, the above three comprehensive proposed models are evaluated under Wilcoxon rank test for better sensitive information. Further, the less object detection loss (2 %) and significant computational time variation (69 %) indicates the performance of the comprehensive helmet violator detection system was good compared to existing study.

Object and Event Detection Pipeline for Rink Hockey Games

All types of sports are potential application scenarios for automatic and real-time visual object and event detection. In rink hockey, the popular roller skate variant of team hockey, it is of great interest to automatically track player movements, positions, and sticks, and also to make other judgments, such as being able to locate the ball. In this work, we present a real-time pipeline consisting of an object detection model specifically designed for rink hockey games, followed by a knowledge-based event detection module. Even in the presence of occlusions and fast movements, our deep learning object detection model effectively identifies and tracks important visual elements in real time, such as: ball, players, sticks, referees, crowd, goalkeeper, and goal. Using a curated dataset consisting of a collection of rink hockey videos containing 2525 annotated frames, we trained and evaluated the algorithm’s performance and compared it to state-of-the-art object detection techniques. Our object detection model, based on YOLOv7, presents a global accuracy of 80% and, according to our results, good performance in terms of accuracy and speed, making it a good choice for rink hockey applications. In our initial tests, the event detection module successfully detected an important event type in rink hockey games, namely, the occurrence of penalties.

Drone-based photogrammetry combined with deep learning to estimate hail size distributions and melting of hail on the ground

Abstract. Hail is a major threat associated with severe thunderstorms, and estimating the hail size is important for issuing warnings to the public. For the validation of existing operational, radar-derived hail estimates, ground-based observations are necessary. Automatic hail sensors, for example within the Swiss Hail Network, record the kinetic energy of hailstones to estimate the hail sizes. Due to the small size of the observational area of these sensors (0.2 m2), the full hail size distribution (HSD) cannot be retrieved. To address this issue, we apply a state-of-the-art custom trained deep learning object detection model to drone-based aerial photogrammetric data to identify hailstones and estimate the HSD. Photogrammetric data of hail on the ground were collected for one supercell thunderstorm crossing central Switzerland from southwest to northeast in the afternoon of 20 June 2021. The hail swath of this intense right-moving supercell was intercepted a few minutes after the passage at a soccer field near Entlebuch (canton of Lucerne, Switzerland) and aerial images were taken by a commercial DJI drone, equipped with a 45-megapixel full-frame camera system. The resulting images have a ground sampling distance (GSD) of 1.5 mm per pixel, defined by the focal length of 35 mm of the camera and a flight altitude of 12 m above the ground. A 2-dimensional orthomosaic model of the survey area (750.4 m2) is created based on 116 captured images during the first drone mapping flight. Hail is then detected using a region-based convolutional neural network (Mask R-CNN). We first characterize the hail sizes based on the individual hail segmentation masks resulting from the model detections and investigate the performance using manual hail annotations by experts to generate validation and test data sets. The final HSD, composed of 18 207 hailstones, is compared with nearby automatic hail sensor observations, the operational weather-radar-based hail product MESHS (Maximum Expected Severe Hail Size) and crowdsourced hail reports. Based on the retrieved data set, a statistical assessment of sampling errors of hail sensors is carried out. Furthermore, five repetitions of the drone-based photogrammetry mission within 18.65 min facilitate investigations into the hail-melting process on the ground.

Detecting Moving Wildlife Using the Time Difference between Two Thermal Airborne Images

Wildlife damage to agriculture is serious in Japan; therefore, it is important to understand changes in wildlife population sizes. Although several studies have been conducted to detect wildlife from drone images, behavioral changes (such as wildlife escaping when a drone approaches) have been confirmed. To date, the use of visible and near-infrared images has been limited to the daytime because many large mammals, such as sika deer (Cervus nippon), are crepuscular. However, it is difficult to detect wildlife in the thermal images of urban areas that are not open and contain various heat spots. To address this issue, a method was developed in a previous study to detect moving wildlife using pairs of time-difference thermal images. However, the user’s accuracy was low. In the current study, two methods are proposed for extracting moving wildlife using pairs of airborne thermal images and deep learning models. The first method was to judge grid areas with wildlife using a deep learning classification model. The second method detected each wildlife species using a deep learning object detection model. The proposed methods were then applied to pairs of airborne thermal images. The classification test accuracies of “with deer” and “without deer” were >85% and >95%, respectively. The average precision of detection, precision, and recall were >85%. This indicates that the proposed methods are practically accurate for monitoring changes in wildlife populations and can reduce the person-hours required to monitor a large number of thermal remote-sensing images. Therefore, efforts should be made to put these materials to practical use.

Tongue feature dataset construction and real-time detection.

Tongue diagnosis in traditional Chinese medicine (TCM) provides clinically important, objective evidence from direct observation of specific features that assist with diagnosis. However, the current interpretation of tongue features requires a significant amount of manpower and time. TCM physicians may have different interpretations of features displayed by the same tongue. An automated interpretation system that interprets tongue features would expedite the interpretation process and yield more consistent results. This study applied deep learning visualization to tongue diagnosis. After collecting tongue images and corresponding interpretation reports by TCM physicians in a single teaching hospital, various tongue features such as fissures, tooth marks, and different types of coatings were annotated manually with rectangles. These annotated data and images were used to train a deep learning object detection model. Upon completion of training, the position of each tongue feature was dynamically marked. A large high-quality manually annotated tongue feature dataset was constructed and analyzed. A detection model was trained with average precision (AP) 47.67%, 58.94%, 71.25% and 59.78% for fissures, tooth marks, thick and yellow coatings, respectively. At over 40 frames per second on a NVIDIA GeForce GTX 1060, the model was capable of detecting tongue features from any viewpoint in real time. This study constructed a tongue feature dataset and trained a deep learning object detection model to locate tongue features in real time. The model provided interpretability and intuitiveness that are often lacking in general neural network models and implies good feasibility for clinical application.

Logistics box recognition in robotic industrial de-palletising procedure with systematic RGB-D image processing supported by multiple deep learning methods

In an automated box de-palletisation system that utilises robots, vision-based box recognition on the pallet plays the main role in providing picking guidelines. The complexity of the working condition and the target object, particularly the cluttered arrangement and various outer surfaces of the boxes, significantly affect the quality of the outcome. Typically, a large-scale vision dataset is required to train a deep learning object-detection model. However, considerable effort and time is required to achieve this. Therefore, this study proposes a Mask R-CNN-based detection approach for box objects, which is supported by a cycle generative adversarial network (Cycle GAN). The purpose of the Cycle-GAN is to optimise the outer surfaces of boxes by automatically erasing tags, stickers, labels, and symbols that exist on the boxes before loading them to the Mask R-CNN for detection. Subsequently, the obtained result was combined with the output from the developed boundary-enhancing technique that was applied to a depth map. Consequently, the box detection performance was significantly improved, and it was confirmed through experiments with a practical robot system in picking tasks. In the experiments, the success rate of the proposed method was validated using 200 cases of orderly and disorderly arrangements of boxes, respectively. Furthermore, the metric of the mean absolute error between the predicted picking point and the ground truth values for the test cases in the implementation process for the robot operation was also researched.

Tropical Cyclone Detection from the Thermal Infrared Sensor IASI Data Using the Deep Learning Model YOLOv3

Tropical cyclone (TC) detection is essential to mitigate natural disasters, as TCs can cause significant damage to life, infrastructure and economy. In this study, we applied the deep learning object detection model YOLOv3 to detect TCs in the North Atlantic Basin, using data from the Thermal InfraRed (TIR) Atmospheric Sounding Interferometer (IASI) onboard the Metop satellites. IASI measures the outgoing TIR radiation of the Earth-Atmosphere. For the first time, we provide a proof of concept of the possibility of constructing images required by YOLOv3 from a TIR remote sensor that is not an imager. We constructed a dataset by selecting 50 IASI radiance channels and using them to create images, which we labeled by constructing bounding boxes around TCs using the hurricane database HURDAT2. We trained the YOLOv3 on two settings, first with three “best” selected channels, then using an autoencoder to exploit all 50 channels. We assessed its performance with the Average Precision (AP) metric at two different intersection over union (IoU) thresholds (0.1 and 0.5). The model achieved promising results with AP at IoU threshold 0.1 of 78.31%. Lower performance was achieved with IoU threshold 0.5 (31.05%), showing the model lacks precision regarding the size and position of the predicted boxes. Despite that, we show YOLOv3 demonstrates great potential for TC detection using TIR instruments data.

Detecting pediatric wrist fractures using deep-learning-based object detection.

Missed fractures are the leading cause of diagnostic error in the emergency department, and fractures of pediatric bones, particularly subtle wrist fractures, can be misidentified because of their varying characteristics and responses to injury. This study evaluated the utility of an object detection deep learning framework for classifying pediatric wrist fractures as positive or negative for fracture, including subtle buckle fractures of the distal radius, and evaluated the performance of this algorithm as augmentation to trainee radiograph interpretation. We obtained 395 posteroanterior wrist radiographs from unique pediatric patients (65% positive for fracture, 30% positive for distal radial buckle fracture) and divided them into train (n = 229), tune (n = 41) and test (n = 125) sets. We trained a Faster R-CNN (region-based convolutional neural network) deep learning object-detection model. Two pediatric and two radiology residents evaluated radiographs initially without the artificial intelligence (AI) assistance, and then subsequently with access to the bounding box generated by the Faster R-CNN model. The Faster R-CNN model demonstrated an area under the curve (AUC) of 0.92 (95% confidence interval [CI] 0.87-0.97), accuracy of 88% (n = 110/125; 95% CI 81-93%), sensitivity of 88% (n = 70/80; 95% CI 78-94%) and specificity of 89% (n = 40/45, 95% CI 76-96%) in identifying any fracture and identified 90% of buckle fractures (n = 35/39, 95% CI 76-97%). Access to Faster R-CNN model predictions significantly improved average resident accuracy from 80 to 93% in detecting any fracture (P < 0.001) and from 69 to 92% in detecting buckle fracture (P < 0.001). After accessing AI predictions, residents significantly outperformed AI in cases of disagreement (73% resident correct vs. 27% AI, P = 0.002). An object-detection-based deep learning approach trained with only a few hundred examples identified radiographs containing pediatric wrist fractures with high accuracy. Access to model predictions significantly improved resident accuracy in diagnosing these fractures.

Intelligent Deployment Solution for Tabling Adapting Deep Learning

The deep learning object detection model can well extract the coordinate information of the separating point of the concentrate ore belt boundary line. However, with the drawbacks of acquired zonal image features, it is difficult to obtain the perfect prediction of the concentrate grade and recovery rate. In this study, by adapting deep learning semantic segmentation technique with DeepLab V3+, which can effectively extract multi-dimensional image features. The mapping relationship between image features and attributes of the Tabling equipment is achieved by constructing a multi-output support vector regression model optimized of a sparrow search algorithm (SSA-MSVR). The beneficiation indicators and operating parameters of the Tabling can be continually detected and optimized by an intelligent system that mainly includes image recognition softwares, automatic control units, data processing and communication workstations, and matching intelligent equipment, which accomplished the intelligent deployment solution of the Tabling beneficiation process.

An automatic inspection system for pest detection in granaries using YOLOv4

Stored-grain pests cause serious economic losses during grain storage. Therefore, it is important for us to know the accurate number and species of stored grain pests as soon as possible so that appropriate measures can be taken to reduce economic losses. However, current research on grain pest detection has two problems. The background of the dataset does not contain any grain, so the results cannot be applied to detect pests in actual situations. The other problem is that methods based on pest traps cannot reflect situations on the surface of the whole granary. This paper proposes an automatic system of pest detection and counting to solve these problems. This system consists of two main parts: a deep learning object detection model called YOLOv4 and a small car with a camera and a supplementary light. And YOLOv4 model that has been trained is embedded in the car. The car can run on the surface of the granary. The camera on the small car can take photos of the pests for YOLOv4 to identify the species and number of pests. In the experiment, a few hundred kilograms of wheat were used to lay a simulated granary. And 2 typical stored-wheat pests, namely, the red flour beetle and the rice weevil, were taken as the research objects. The experimental results demonstrated that the mean average precision (mAP) of the proposed method reached 97.55%, which can meet the accuracy requirements of the detection and counting of pests in the granary in practical applications. The system solves the randomness and insufficient accuracy of pest traps and human eye recognition. And the system can be used to the early warning of pests in granary, which has high accuracy and completely release manual labor.

A feasibility study on the development and use of a deep learning model to automate real-time monitoring of tumor position and assessment of interfraction fiducial marker migration in prostate radiotherapy patients * *Deep Learning Fiducial Marker Localisation.

Purpose. The aim of this study was to assess the feasibility of the development and training of a deep learning object detection model for automating the assessment of fiducial marker migration and tracking of the prostate in radiotherapy patients. Methods and Materials. A fiducial marker detection model was trained on the YOLO v2 detection framework using approximately 20,000 pelvis kV projection images with fiducial markers labelled. The ability of the trained model to detect marker positions was validated by tracking the motion of markers in a respiratory phantom and comparing detection data with the expected displacement from a reference position. Marker migration was then assessed in 14 prostate radiotherapy patients using the detector for comparison with previously conducted studies. This was done by determining variations in intermarker distance between the first and subsequent fractions in each patient. Results. On completion of training, a detection model was developed that operated at a 96% detection efficacy and with a root mean square error of 0.3 pixels. By determining the displacement from a reference position in a respiratory phantom, experimentally and with the detector it was found that the detector was able to compute displacements with a mean accuracy of 97.8% when compared to the actual values. Interfraction marker migration was measured in 14 patients and the average and maximum standard deviation marker migration were found to be and respectively. Conclusion. This study demonstrates the benefits of pairing deep learning object detection, and image-guided radiotherapy and how a workflow to automate the assessment of organ motion and seed migration during prostate radiotherapy can be developed. The high detection efficacy and low error make evident the advantages of using a pre-trained model to automate the assessment of the target volume positional variation and the migration of fiducial markers between fractions.

Remote Sensing Image Retrieval Based on Key Region Detection

Remote sensing (RS) images usually describe large-scale natural geographical scenes with complex and rich background information, which will affect the retrieval performance of image features. How to reduce the background interference and improve the reliability of remote sensing image retrieval (RSIR) features is a problem that needs to be solved. In this paper, a RSIR method based on key region detection was proposed. Firstly, the ground objects of the image are extracted by a famous deep learning object detection model, a YOLO v5 model. Next, we extract the key region of the image according to these ground objects. Then, the image content in the key region is used to extract the retrieval feature by the convolutional neural networks (CNN) model, Resnet. Moreover, the weighted distance based on class probability is used to further improve retrieval performance. Our method utilizes the object detection capability of the YOLO model and the feature extraction capability of RESNET. Our method uses the target detection ability of the YOLO model and the feature extraction ability of RESNET to extract the retrieval feature of RS images. The experimental results on UCMD show that this method can improve the performance of RSIR.

Improved Method to Detect the Tailings Ponds from Multispectral Remote Sensing Images Based on Faster R-CNN and Transfer Learning

The breaching of tailings pond dams may lead to casualties and environmental pollution; therefore, timely and accurate monitoring is an essential aspect of managing such structures and preventing accidents. Remote sensing technology is suitable for the regular extraction and monitoring of tailings pond information. However, traditional remote sensing is inefficient and unsuitable for the frequent extraction of large volumes of highly precise information. Object detection, based on deep learning, provides a solution to this problem. Most remote sensing imagery applications for tailings pond object detection using deep learning are based on computer vision, utilizing the true-color triple-band data of high spatial resolution imagery for information extraction. The advantage of remote sensing image data is their greater number of spectral bands (more than three), providing more abundant spectral information. There is a lack of research on fully harnessing multispectral band information to improve the detection precision of tailings ponds. Accordingly, using a sample dataset of tailings pond satellite images from the Gaofen-1 high-resolution Earth observation satellite, we improved the Faster R-CNN deep learning object detection model by increasing the inputs from three true-color bands to four multispectral bands. Moreover, we used the attention mechanism to recalibrate the input contributions. Subsequently, we used a step-by-step transfer learning method to improve and gradually train our model. The improved model could fully utilize the near-infrared (NIR) band information of the images to improve the precision of tailings pond detection. Compared with that of the three true-color band input models, the tailings pond detection average precision (AP) and recall notably improved in our model, with the AP increasing from 82.3% to 85.9% and recall increasing from 65.4% to 71.9%. This research could serve as a reference for using multispectral band information from remote sensing images in the construction and application of deep learning models.

Deep-learning single-shot detector for automatic detection of brain metastases with the combined use of contrast-enhanced and non-enhanced computed tomography images

PurposeTo develop a deep-learning object detection model for automatic detection of brain metastases that simultaneously uses contrast-enhanced and non-enhanced images as inputs, and to compare its performance with that of a model that uses only contrast-enhanced images. MethodA total of 116 computed tomography (CT) scans of 116 patients with brain metastases were included in this study. They showed a total of 659 metastases, 428 of which were used for training and validation (mean size, 11.3 ± 9.9 mm) and 231 were used for testing (mean size, 9.0 ± 7.0 mm). Single-shot detector (SSD) models were constructed with a feature fusion module, and their results were compared per lesion at a confidence threshold of 50%. ResultsThe sensitivity was 88.7% for the model that used both contrast-enhanced and non-enhanced CT images (the CE + NECT model) and 87.6% for the model that used only contrast-enhanced CT images (the CECT model). The positive predictive value (PPV) was 44.0% for the CE + NECT model and 37.2% for the CECT model. The number of false positives per patient was 9.9 for the CE + NECT model and 13.6 for the CECT model. The CE + NECT model had a significantly higher PPV (t test, p < 0.001), significantly fewer false positives (t test, p < 0.001), and a tendency to be more sensitive (t test, p = 0.14). ConclusionsThe results indicate that the information on true contrast enhancement obtained by comparing the contrast-enhanced and non-enhanced images may prevent the detection of pseudolesions, suppress false positives, and improve the performance of deep-learning object detection models.

A cost-effective maize ear phenotyping platform enables rapid categorization and quantification of kernels.

High-throughput phenotyping systems are powerful, dramatically changing our ability to document, measure, and detect biological phenomena. Here, we describe a cost-effective combination of a custom-built imaging platform and deep-learning-based computer vision pipeline. A minimal version of the maize (Zea mays) ear scanner was built with low-cost and readily available parts. The scanner rotates a maize ear while a digital camera captures a video of the surface of the ear, which is then digitally flattened into a two-dimensional projection. Segregating GFP and anthocyanin kernel phenotypes are clearly distinguishable in ear projections and can be manually annotated and analyzed using image analysis software. Increased throughput was attained by designing and implementing an automated kernel counting system using transfer learning and a deep learning object detection model. The computer vision model was able to rapidly assess over 390 000 kernels, identifying male-specific transmission defects across a wide range of GFP-marked mutant alleles. This includes a previously undescribed defect putatively associated with mutation of Zm00001d002824, a gene predicted to encode a vacuolar processing enzyme. Thus, by using this system, the quantification of transmission data and other ear and kernel phenotypes can be accelerated and scaled to generate large datasets for robust analyses.

Accuracy of a Smartphone-Based Object Detection Model, PlantVillage Nuru, in Identifying the Foliar Symptoms of the Viral Diseases of Cassava-CMD and CBSD.

Nuru is a deep learning object detection model for diagnosing plant diseases and pests developed as a public good by PlantVillage (Penn State University), FAO, IITA, CIMMYT, and others. It provides a simple, inexpensive and robust means of conducting in-field diagnosis without requiring an internet connection. Diagnostic tools that do not require the internet are critical for rural settings, especially in Africa where internet penetration is very low. An investigation was conducted in East Africa to evaluate the effectiveness of Nuru as a diagnostic tool by comparing the ability of Nuru, cassava experts (researchers trained on cassava pests and diseases), agricultural extension officers and farmers to correctly identify symptoms of cassava mosaic disease (CMD), cassava brown streak disease (CBSD) and the damage caused by cassava green mites (CGM). The diagnosis capability of Nuru and that of the assessed individuals was determined by inspecting cassava plants and by using the cassava symptom recognition assessment tool (CaSRAT) to score images of cassava leaves, based on the symptoms present. Nuru could diagnose symptoms of cassava diseases at a higher accuracy (65% in 2020) than the agricultural extension agents (40–58%) and farmers (18–31%). Nuru’s accuracy in diagnosing cassava disease and pest symptoms, in the field, was enhanced significantly by increasing the number of leaves assessed to six leaves per plant (74–88%). Two weeks of Nuru practical use provided a slight increase in the diagnostic skill of extension workers, suggesting that a longer duration of field experience with Nuru might result in significant improvements. Overall, these findings suggest that Nuru can be an effective tool for in-field diagnosis of cassava diseases and has the potential to be a quick and cost-effective means of disseminating knowledge from researchers to agricultural extension agents and farmers, particularly on the identification of disease symptoms and their management practices.

Vehicle Detection and Ranging Using Two Different Focal Length Cameras

Vehicle detection is a crucial task for autonomous driving and demands high accuracy and real-time speed. Considering that the current deep learning object detection model size is too large to be deployed on the vehicle, this paper introduces the lightweight network to modify the feature extraction layer of YOLOv3 and improve the remaining convolution structure, and the improved Lightweight YOLO network reduces the number of network parameters to a quarter. Then, the license plate is detected to calculate the actual vehicle width and the distance between the vehicles is estimated by the width. This paper proposes a detection and ranging fusion method based on two different focal length cameras to solve the problem of difficult detection and low accuracy caused by a small license plate when the distance is far away. The experimental results show that the average precision and recall of the Lightweight YOLO trained on the self-built dataset is 4.43% and 3.54% lower than YOLOv3, respectively, but the computing speed of the network decreases 49 ms per frame. The road experiments in different scenes also show that the long and short focal length camera fusion ranging method dramatically improves the accuracy and stability of ranging. The mean error of ranging results is less than 4%, and the range of stable ranging can reach 100 m. The proposed method can realize real-time vehicle detection and ranging on the on-board embedded platform Jetson Xavier, which satisfies the requirements of automatic driving environment perception.

A Power-Efficient Optimizing Framework FPGA Accelerator Based on Winograd for YOLO

Accelerating deep learning networks in edge computing based on power-efficient and highly parallel FPGA platforms is an important goal. Combined with deep learning theory, an accelerator design method based on the Winograd algorithm for the deep learning object detection model YOLO under the PYNQ architecture is proposed. A Zynq FPGA is used to build the hardware acceleration platform of a YOLO network. The Winograd algorithm is used to improve traditional convolution. In the FPGA, the numerous multiplication operations in the YOLO network are converted into addition operations, reducing the computational complexity of the model. The data of the original model are processed at a low fixed point, reducing the resource consumption of the FPGA. To optimize memory, a buffer pipeline method is proposed, which further improves the efficiency of the designed accelerator. Experiments show that compared with the acceleration of the YOLO model based on GPUs and other FPGA platforms, the proposed method not only optimizes FPGA resource usage but also reduces power consumption to 2.7 W. Additionally, the detection accuracy loss is less than 3%.

Hierarchical Image Object Search Based on Deep Reinforcement Learning

Abstract Object detection technology occupies a pivotal position in the field of modern computer vision research, its purpose is to accurately locate the object human beings are looking for in the image and classify the object. With the development of deep learning technology, convolutional neural networks are widely used because of their outstanding performance in feature extraction, which greatly improves the speed and accuracy of object detection. In recent years, reinforcement learning technology has emerged in the field of artificial intelligence, showing excellent decision-making ability to deal with problems. In order to combine the perception ability of deep learning technology with the decision-making ability of reinforcement learning technology, this paper incorporate reinforcement learning into the convolutional neural network, and propose a hierarchical deep reinforcement learning object detection model.