Detection of maize tassels for UAV remote sensing image with an improved YOLOX Model

SONG Chao-yu ,ZHANG Fan ,LI Jian-sheng ,XIE Jin-yi ,YANG Chen ,ZHOU Hang ,ZHANG Junxiong#

1 College of Engineering, China Agricultural University, Beijing 100083, P.R.China

2 College of Agronomy and Biotechnology, China Agricultural University, Beijing 100083, P.R.China

Abstract Maize tassel detection is essential for future agronomic management in maize planting and breeding,with application in yield estimation,growth monitoring,intelligent picking,and disease detection.However,detecting maize tassels in the field poses prominent challenges as they are often obscured by widespread occlusions and differ in size and morphological color at different growth stages.This study proposes the SEYOLOX-tiny Model that more accurately and robustly detects maize tassels in the field.Firstly,the data acquisition method ensures the balance between the image quality and image acquisition efficiency and obtains maize tassel images from different periods to enrich the dataset by unmanned aerial vehicle (UAV).Moreover,the robust detection network extends YOLOX by embedding an attention mechanism to realize the extraction of critical features and suppressing the noise caused by adverse factors(e.g.,occlusions and overlaps),which could be more suitable and robust for operation in complex natural environments.Experimental results verify the research hypothesis and show a mean average precision (mAP@0.5) of 95.0%.The mAP@0.5,mAP@0.5–0.95,mAP@0.5–0.95(area=small),and mAP@0.5–0.95(area=medium) average values increased by 1.5,1.8,5.3,and 1.7%,respectively,compared to the original model.The proposed method can effectively meet the precision and robustness requirements of the vision system in maize tassel detection.

Keywords: maize,tassel detection,remote sensing,deep learning,attention mechanism

1.Introduction

Maize is a significant crop for food consumption,feed raw materials,and a source of material for a wide range of industrial products and plays an essential part in agricultural production and the national economy (Ranumet al.2014;Suet al.2019).Most maize cultivation fields are planted with hybrid maize.However,the advantages of hybrids maize decrease after the second generation,and self-pollination is not desired (Kurtulmu? and Kavdir 2014;Jiet al.2021).A major challenge in maize breeding is to ensure seed purity and maximize maize yield in the context of rising seed farming costs and skilled labor shortages.To ensure the production of only hybrid seed,it is essential to constantly monitor production fields,with particular attention given to maize tassel recognition during breeding.Therefore,rapid,accurate,and low-cost identification of maize tassels during the maize flowering stage is significant for maize breeding.

With the development of computer vision,an increasing number of agricultural technologies based on computer vision are being developed to reduce labor costs and improve efficiency.These include crop growth stage prediction (Yeet al.2013),selection seeds (Liet al.2018),leaf disease detection (Camargo and Smith 2009),crop biomass monitoring (Jannouraet al.2015),and fruit development monitoring (Changet al.2018).Meanwhile,efforts have also been made to detect maize tassels and maize tassel segmentation by computer vision.Firstly,researchers identify maize tassels by support vector machines based on the color information of maize tassels (Kurtulmu? and Kavdir 2014).Luet al.(2015)further extended the range of characterizing the maize tassel trait with computer vision technology.However,with the deepening of study contexts for the detection of maize tassels,some problems that affect the detection accuracy are becoming more prominent,such as complex background,illumination,soil color,shadow,shape,and size of tassels (Yeet al.2013;Gageet al.2017;Liuet al.2020;Karamiet al.2021).Moreover,the sensor that obtains the information on maize tassels is low-throughput,the efficiency is relatively low,and it is unlikely to be applied to the monitoring and detection of maize tassels in the field.

Technological advances in the development of unmanned aerial vehicles (UAVs) equipped with sensors are making a significant impact in field-based crop phenotyping (Zhang and Kovacs 2012;Bendiget al.2015;Jinet al.2017;Kohet al.2019).Moreover,the improved computer performance and deep learning models demonstrated outstanding performance for many precision agriculture applications (Oscoet al.2021;Panet al.2022;Yanget al.2022).Liuet al.(2020) used a faster region-based convolutional neural network (Renet al.2017) to detect maize tassels for UAV images.The detection accuracy can reach 89.96%.Yanget al.(2021)proposed an improved CenterNet that embeds location information in the feature extraction module to improve the detection accuracy,reaching 92.4%.However,the above methods for tassels are aimed at single tassels of a specific species in a specific environment and cannot distinguish the tassels in the early tasseling stage (Zanet al.2020),which is challenging to meet the demand for dynamic monitoring of maize tassels for maize breeding.Zanet al.(2020) proposed a detection method for maize tassels in different tasseling stages by combining Random Forest Classifier and the VGG16 network.However,this method is easily affected by occluded and small objects for maize tassels,and the data collection time is much shorter than the flowering period of maize tassels,which causes the model to lack robustness.Although many researchers have made significant progress in this field in various ways,the above methods still lack robustness and accuracy for detecting small maize tassels,which cannot meet real-world application needs.

Motivated by the abovementioned challenges,a maize tassel detection method of UAV remote sensing image based on an improved YOLOX model is proposed (Geet al.2021).Firstly,the data acquisition method can better balance image quality and image acquisition efficiency.It can obtain maize tassel images from different periods to enrich our dataset by UAV and ensure that the model proposed in this study detects the tassels of different varieties and at different tasseling stages.Then,the Squeeze-and-Excitation (SE) attention mechanism (Huet al.2020) is embedded into YOLOX to focus on the informative pixels and suppress the noise.Finally,Varifocal Loss (Zhanget al.2021) is employed to reduce the influence of the foreground-background imbalance problem in YOLOX.Experimental results in the real-world setting show that the proposed method performs well in detecting maize tassels,achieving higher accuracy and stronger robustness.These findings suggest that the proposed method may be more suitable for real-world applications to detect maize tassels during the flowering process.

The rest of this paper is organized as follows: Section 2 presents the datasets,image acquisition method,and the proposed SEYOLOX-tiny framework;next,in Section 3,the experiment validates that the method outperforms others in terms of precision and robustness,it also discusses the results of the test phases;finally,Section 4 concludes the SEYOLOX-tiny Model’s characteristics and clarifies the other unsolved problems in this field,providing future research directions.

2.Materials and methods

2.1.Image acquisition

Study areaThe experiment was carried out at the Shangzhuang Experiment Station of China Agricultural University (40°06′5′′N,116°12′22′′E,Beijing,China)during the 2021 growing season (Fig.1-A).The trial area was approximately 56.5 m from north to south and 14 m from east to west.The spring maize planted is ripe for the year,sown in mid-April 2021,and harvested in early August.Moreover,the row spacing was 0.6 m,and the plant spacing was 0.3 m.

Fig.1 Image sampling.A,the field trials at the Shangzhuang Experiment Station,Beijing,China.B,work scenario.C,the UAV DJI Phantom 4 Multispectral and the interpretation of the gimbal angle.

Image collection platformThe UAV DJI Phantom 4 Multispectral (P4M) was used for experimental data collection.The P4M’s imaging system contains six cameras (resolution: 1 630×1 300 pixels) with 1/2.9” CMOS sensors,including an RGB camera (P4 Multispectral,DJI,China),at 2MP with global shutter,on a 3-axis stabilized gimbal (Fig.1-C).This experiment used visible-light images to detect maize tassels at the tasseling stages.

Balance flying height,angle,and image qualityDue to the growth characteristics of maize tassels,the exposed parts of the tassels in the early tasseling stage are tiny,coupled with the tassel easily obscured by leaves,significantly challenging the detection of maize tassels.In order to solve the above problems,UAVs usually use high-pixel cameras and then enlarge the object through post-processing or lower the flight altitude to reduce the number of small objects in the image.The latter may become a better choice under the development trend of pursuing efficient and lightweight neural networks.However,the problem with this is that when the UAV takes images in the field,the lifting force generated by the rotation of its blades will also react to the crops below the UAV.Suppose the flying height of the UAV is too low,in that case,the images containing the object information can be drastically reduced,and more seriously,this can cause damage to crops and the UAV itself.

Image acquisitionIn this paper,the image acquisition method considers the influence of the gimbal angle on the acquisition of image data.The gimbal angles are set to 30°,40°,50°,and 60°,respectively,and the UAV is lowered to a suitable height to ensure that the maize plants are not damaged (Fig.2).The UAV takes images at different heights and angles and compares the number of maize tassels in the image and the quality of maize tassels.It was finally determined that the flight height was 7 m,and the angle of the gimbal was 50°.

Fig.2 The UAV obtain pictures of maize tassels under the change of the gimbal angle.A–E,angle=30°,40°,50°,60°,and 60°,respectively.

To ensure that the model proposed in this study can detect maize tassels of different sizes and tasseling stages,we collected eight data groups with a range of 8 days from 15–22 July 2021.Maize tassels’ shapes in different periods are different,and the shapes of maize tassels in the same stage are also different (Fig.3-D).In order to prevent excessive duplication of sample data,the obtained data set was sampled every two times,and the images with poor quality in the samples due to lens distortion were eliminated.Finally,456 images of maize tassels were obtained.

Fig.3 The size of maize tassels in different tasseling stages.A,the early tasseling stage.B,the middle tasseling stage.C,the late tasseling stage.D,unsynchronized growth stage.E,the varied size of tassels in the tasseling stage.

2.2.Image preprocessing and dataset analysis

Image preprocessingAs shown in Fig.3-E,it was evident that the size of maize tassels is relatively tiny,and the occlusion is widespread.Suppose the original image is used in this experiment directly.In that case,the feature extraction process of the YOLOX network will over-compress the tiny object of the maize tassel and affect the detection speed of the maize tassel.In order to weaken the impact of the above phenomenon,two datasets were proposed.

(1) The first dataset: The original images were cropped and filtered into 2 736 images with 640×640 pixels resolution as the first dataset (Fig.4).The training,validation,and test sets had 2 188,274,and 274 images in the ratio of 8:1:1,respectively.

Fig.4 Cutting image.

(2) The second dataset: The randomly selected 120 images from the original images of maize tassels were obtained as the second dataset.In order to more reliably test the network performance,the second dataset of 120 images was randomly divided into three groups of 40 images,Test 1,Test 2,and Test 3.

A rigorous tassel annotation was performed for this study using the open-source annotation tool Labelme.All the pixels of maize tassels were within the range of the bounding boxes.

Dataset analysisConsulting our data labeling information,we adopted common object size classification standards,identifying the objects with dimensions smaller than 32×32 pixels as small objects,those larger than 96×96 pixels as large objects,and the rest as medium objects.The guide modification was made based on the YOLOX network by analyzing specific size information and distributions of maize tassels.

(1) The first dataset: From the statistical results,the proportion of small objects and large objects is 2.13 and 8.97%,respectively,and the proportion of medium objects is 88.9% in the first dataset.

(2) The second dataset: We ultimately wanted the network to be tested on the second dataset and achieve optimal results.However,the input and test sizes were set to 640×640 pixels of our network model in this paper.The images in the second dataset need to be scaled from 1 630×1 300 pixels to 640×640 pixels.Therefore,in the second dataset,analyzing those scaled drawings by the object size classification standards,the proportion of small,medium,and large objects are 62.46,37.47,and 0.07%,respectively.

The specific data of object size distribution and classification proportion of those datasets are shown in Fig.5.

Fig.5 Size distribution of maize tassels.A,size distribution of maize tassels of the first dataset.B,classification proportion of maize tassels of the first dataset.C,classification proportion of maize tassels of the second dataset.

2.3.Training environment and evaluation indicators

The framework for deep learning was PyTorch1.9.0.The experiment was run on Windows 10,with an AMD Ryzen 75 800H with Radeon Graphics CPU at basic frequency 3 201 MHz,16 GB RAM,an Nvidia GeForce RTX 3060 Laptop GPU,and Compute Unified Device Architecture (CUDA) 11.3,accelerated by cuDNN version 8.2.1.We used a learning rate set to learning divided by 128,with an initial learning set of 0.01,and the cosine learning schedule.The weight decay is 0.0005,and the SGD momentum is 0.9.The batch sizes were eight by default to single GPU devices.The input and test image’s resolution size was set at 640×640 pixels,and the number of iterations was set to 100.

The evaluation indicators of that the model of YOLOX compared with other models were mainly based on precision,recall,and mean average precision (mAP),respectively:

where TP,FP,and FN are the number of true positives,false positives,and false negatives,respectively.N is the number of all maize tassels detected by the network model.In eq.(4),AP is the area under the precisionrecall curve (P-R curve),and mAP is the average of AP for different categories.The data only had one maize tassel category in our experiment,and thus,nwas 1.Where mAP@0.5represents the mean value of mAP when the IOU threshold is 0.5,where mAP@0.5–0.95represents the mean value of mAP at different IOU thresholds (from 0.5 to 0.95,step-size 0.05);mAP@0.5–0.95(area=small)represents mAP@0.5–0.95for the small objects;mAP@0.5–0.95(area=medium)represents mAP@0.5–0.95for the medium objects.

2.4.SEYOLOX-tiny for object detection

YOLOX is a new high-performance detector of the YOLO series,and its network architecture baseline adopts the architecture of DarkNet53 and YOLOv3.This combination has experienced improvements,such as anchor-free,a decoupled head,and the leading label assignment strategy Simplified Optimal Transport Assignment (SimOTA) (Geet al.2021),to achieve state-of-the-art results.It won first place in the streaming Perception Challenge (Workshop on Autonomous Driving at CVPR 2021).

YOLOX provides a large-scale range of models,including YOLOX-Nano,YOLOX-tiny,YOLOX-s,YOLOX-m,YOLOX-l,and YOLOX-x,the parameters and GFLOPs of which are gradually increasing.These models were trained separately,and the performance of those detectors is listed in Table 1.As shown in Table 1,the mAP@0.5–0.95of the YOLOX-tiny Model increases by 1.92% compared with that YOLOX-Nano model;it is only 0.86 and 1.07% less than the mAP@0.5–0.95of the YOLOX-s and YOLOX-m models,respectively.Moreover,the parameters of the YOLOX-tiny Model are smaller than the YOLOX-s and YOLOX-m models by 4.94 and 20.24 M,respectively,but only 4.15 M larger than the parameters of the YOLOX-Nano Model.Therefore,the YOLOXtiny Network Model was adopted in this experiment after considering the network’s detection accuracy and parameters.

Table 1 Comparison of model prediction results and the counterparts in terms of average precision (AP,%) on COCO val1)

The overall model of the YOLOX network mainly includes four parts: the end of Input,Backbone,Neck,and Head.Although the YOLOX-tiny Model is the best choice for balancing accuracy and size,Table 1 shows that all models are not very good for small object detection,and there are still some details that can be improved in this network.Therefore,we modified the structure of YOLOXtiny,combined the SE module with every Up-sample unit in the Neck to form the SE-Up module,and replaced Varifocal Loss (Zhanget al.2021) with BCE Loss for the training object branch.The structure of the SEYOLOXtiny Model is shown in Fig.6.

Fig.6 The structure of the SEYOLOX-tiny Model.Conv,convolution;BN,batch normalization;Silu,sigmoid linear unit;DWConv,depthwise separable convolution;BaseConv,base convolution.

The SE module allocates limited computing resources to the part with the highest proportion of information in the image.It can better gather the attention of the network model to recognize the object and reduce the influence of irrelevant background.Moreover,considering the establishment of dependencies between channel information in the internal feature map of the module,the global average pooling is injected into the module channel.Then the global information of the feature map is encoded and covered to the original feature map,and the structure of the SE module is shown in Fig.7.

Fig.7 The structure of Squeeze-and-Excitation (SE) module.FC,fully connected layer;ReLu,rectified linear unit.

The end of InputThe input end mainly includes a robust data enhancement module,which adds Mosaic and MixUp into its augmentation strategies to boost the YOLOX’s performance.Mosaic and MixUp are both efficient augmentation strategies and can improve the generalization performance and robustness of the model.

However,as it is better to weaken the augmentation for small models (Geet al.2021),we did not use the MixUp method in our experiment.Mosaic can reduce the number of graphics processing units (GPUs) required and training time.Some images of the robust data enhancement are shown in Fig.8.

Fig.8 Mosaic data enhancement.

BackboneThe Backbone is composed of the structures of Focus,CPS,and Spatial Pyramid Pooling (SPP).The Focus module mainly performs a slicing operation on the image,expands the input channel four times through the slicing operation,and can obtain the double-sampled downsampling feature map without information loss,making the model more compact.There are two CSP structures used in YOLOX-tiny,CSP1_x consists of the Backbone,and CSP2_x consists of the Neck.The module of SPP can generate a fixed-size output,regardless of the input size,and fuse local features with global features,enriching the expressive power of feature maps.

NeckThe Neck module includes the structures of the Feature Pyramid Network (FPN),which primarily helps extract more vital features by fusing Top-level features with low-level features through up-sampling.

HeadThe most distinctive feature of the head of YOLOX is the decoupled head (Fig.9),which splits classification and localization on a fully connected head and a convolution head,respectively.Second,the strategies of SimOTA,an improvement of OTA (Geet al.2021),can reduce cost and achieve a similar effect in the label assignment stage from a global perspective,as compared to OTA.Moreover,unlike before the YOLO series,the anchor-free mechanism is designed in YOLOX,which significantly reduces the number of design parameters that need slight tuning and many tricks involved for good performance.YOLOX uses BCE Loss for the training class and objects branch and IoU Loss for the training reg branch.Since the complex background of the maize tassel image is obtained in the field,maize tassels appear numerous and dense.We have improved the training of the object branch by replacing BCE Loss with the Varifocal Loss.

Fig.9 YOLOX decoupled head.Cls.,classification;Reg.,regression;IoU.,intersection over union.

The Varifocal Loss (Zhanget al.2021) was inspired by the Focal Loss (Linet al.2017),which is designed to reduce the impact of the problem of the extreme imbalance between foreground and background classes during the training of dense object detectors.The Varifocal Loss is defined as:

wherepis the predicted the IoU-aware classification score(IACS),qis the object score,αis an adjustable scaling factor to the negative loss term,andγis an adjustable scaling factor to the losses.

3.Results

3.1.Improvement in the network model training

As mentioned previously,we combined the SE module with every up-sample unit in the Neck to form the SE-Up module that can improve the network model’s accuracy for detecting maize tassels,considering the varied size of maize tassels and the maize growth in complex background.The change in loss function and accuracy of the two network models were compared and analyzed,and the corresponding change curve was drawn in Fig.10.

Fig.10 Comparison of the loss function and accuracy of the two network models.A,comparison of the loss function.The Loss suffix with -Y indicates the Loss generated by the YOLOX Model training process,and -S indicates the Loss generated by the SEYOLOX-tiny Model training process.B,comparison of the accuracy of the first dataset (640×640 pixels resolution).mAP@0.5,mean average precision;mAP@0.5–0.95,the mean value of mAP at different IOU thresholds (from 0.5 to 0.95,step-size 0.05).The mAP suffix with -Y indicates the mean average precision (mAP) generated by the YOLOX Model training process,and -S indicates the mAP generated by the SEYOLOX-tiny Model training process.C,comparison of the accuracy of the second dataset (1 630×1 300 pixels resolution): where Test 1-Y represents the original YOLOX-tiny Model prediction results on the Test1 dataset,Test 1-S represents the improved YOLOX-tiny Model prediction results on the Test 1 dataset,Mean-Y represents the YOLOX-tiny Model average prediction results on Test 1,Test 2,and Test 3 datasets,and Mean-S represents the SEYOLOX-tiny Model average prediction results on the Test 1,Test 2,and Test 3 dataset.

The curve of the loss function of the two network models is shown in Fig.10-A;the value of the four loss functions in the SEYOLOX-tiny Model after training is lower than that in the original YOLOX-tiny Model.In addition,the SEYOLOX-tiny Model has a smaller amplitude of the oscillation in training.A marked increase in the detection accuracy of the SEYOLOX-tiny Model was shown in Fig.10-B.In particular,the mAP@0.5–0.95,mAP@0.5–0.95(area=small),and mAP@0.5–0.95(area=medium)increased by 2.4,6.2,and 1.6%,respectively,which further proved the effectiveness of the SEYOLOX-tiny Model in detecting regions of maize tassels in the first dataset.

3.2.Improvement in detection accuracy in the second dataset

The SEYOLOX-tiny Model meets good results in the second dataset,as shown in Fig.10-C.The results in Fig.10-C show the accuracy of the SEYOLOX-tiny Model that mAP@0.5,mAP@0.5–0.95,mAP@0.5–0.95(area=small),and mAP@0.5–0.95(area=medium)average increased by 1.5,1.8,5.3,and 1.7%,respectively,compared to the original model.The above results prove that the SEYOLOX-tiny model can more accurately distinguish maize tassels in a complex environment.Also,reducing the influence that more upsampling will affect the quality of the feature map and cause problems such as object feature loss and blurring.Some comparison results are shown in Fig.11.

Fig.11 Comparison of the model results.The orange and red boxes represent the original prediction and absent detection,respectively.A,C,and E were detected by the YOLOX-tiny Model;B,D,and F were detected by the SEYOLOX-tiny Model.

As shown in Fig.11,compared with the YOLOXtiny Model,the SEYOLOXtiny Model has a lower rate of missed detection.A total of three images were tested,each representing a tassel state.The first,second,and third images have 17,48,and 65 maize tassels,respectively,for a total of 130 maize tassels.The YOLOXtiny model found 110 maize tassels,and the SEYOLOXtiny found 125 maize tassels in the sum.The proposed method maintains a good detection effect of maize tassels,which indicates that the proposed model is robust in a complex environment.

3.3.Comparison of different attention mechanisms for YOLOX-tiny

In order to show that the performance of SE was superior to the others in this experiment,we combined the SE,Efficient Channel Attention (ECA) (Wanget al.2020),and Convolutional b l o c k a t t e n t i o n m o d u l e(CBAM) (Wooet al.2018)with the network model in the same position,respectively.All were evaluated on the second dataset.A comparison of the predicted results is presented in Table 2.Compared with the other two models,the model that combined the SE module is more accurate at identifying maize tassels in complex backgrounds.

Table 2 Comparison of model evaluation indicators (average evaluation results on the Test 1,Test 2,and Test 3 datasets)1)

4.Discussion

Detecting maize tassels in natural canopy images poses various difficulties and challenges (Kurtulmu? and Kavdir 2014).Maize tassels in the field differ in size and morphological color at different growth stages.The researchers improved the prediction accuracy to 89.96%for detecting maize tassels from UAV images by using the Faster R-CNN (Liuet al.2020).Yanget al.(2021)proposed an improved CenterNet to improve the detection accuracy,reaching 92.4%.However,the above research did not identify tassels in different tasseling stages,which poses a challenge to meeting the demand for dynamic monitoring of maize tassels in maize breeding.Zanet al.(2020) paid attention to this problem and proposed a detection method for maize tassels in different tasseling stages by combining Random Forest Classifier and the VGG16 network;the detection accuracy reached 90.4%.However,the detection accuracy of maize tassels in the early tasseling stages was low,mainly due to the small size and occlusion phenomenon of maize tassels (Zanet al.2020).These challenges become increasingly significant when detecting maize tassels in the real-world scene.

Therefore,this study exploits the combination of the image acquisition method and maize tassel detection model based on the YOLOX network,aiming to more accurately detect the maize tassels and address the falserecognition problem caused by occlusion and small size.Inconsistent with previous studies of the image acquisition method (Wanget al.2020),we found that balancing flying height and the gimbal angle can reduce the number of small maize tassels in the image and improve the quality of maize tassels.Moreover,embedding the attention mechanism in the structures of the Feature Pyramid Network and replacing Varifocal Loss with BCE Loss for the training object branch can alleviate the influence of tassels’ occlusion and small size on the model’s performance.The experimental results show that the mean average precision (mAP@0.5) was 95.0% for the detection of maize tassels.The YOLOX-tiny is better at detecting maize tassels of various sizes and more suitable for deploying on a vision system for detecting maize tassels.

The visualization results in Fig.11 show that the proposed method performs well in various maize tassels and different tasseling stages.Most small objects can be detected,even including the tiny maize tassels.However,there are still some situations of missed detections,such as Fig.11-D,which clearly shows the individual extremes in the detection of maize tassels,severe occlusion,tiny size,and excessive overlap.Furthermore,severe occlusion and tiny size are widespread in the early tasseling stage in Fig.11-A.Despite some missed detections,this study indicates that improving data acquisition methods and modifying the network for dataset characteristics can achieve good results.

5.Conclusion

In order to effectively detect maize tassels in a complex background during the flowering process,this study used UAV to collect images of maize tassels in different periods to enrich the image dataset.The flight height was 7 m,and the gimbal angle was 50°,which better balances the scope and quality of image acquisition.Meantime we proposed a SEYOLOX-tiny architecture that extends YOLOX by adding an attention mechanism of Squeeze-and-Excitation(SE).Moreover,we proposed replacing Varifocal Loss with BCE Loss for the training object branch to increase intention to those easily overlooked features and reduce the impact of adverse factors such as occlusion,varied sizes,and overlap.The experimental results show that the mean average precision (mAP@0.5) was 95.0%.Moreover,the mAP@0.5,mAP@0.5–0.95,mAP@0.5–0.95(area=small),and mAP@0.5–0.95(area=medium)average increased by 1.5,1.8,5.3,and 1.7%,respectively,compared to the original model.The proposed SEYOLOX-tiny network performs well in detecting maize tassels.It exhibits stronger robustnessand is more suitable for detecting maize tassels during the flowering process in real-world scenes.

Acknowledgements

This research was supported by the Chinese Universities Scientific Fund (2022TC169).Furthermore,the authors are grateful to Shangzhuang Experiment Station of China Agricultural University for the support in performing the experiments.

Declaration of competing interest

The authors declare that they have no conflict of interest.