Introduction

OD in RSI is an essential yet complex endeavor, confronted with numerous obstacles in real-world applications. Over the past years, significant advancements have been witnessed in the study of RSI, leading to data that exhibit extensive multi-scale attributes, substantial object overlays, and non-uniform data dispersion, which brings many difficulties to OD1,2.

Figure 1
figure 1

OD algorithm flow chart.

Full size image

Figure 1 depicts the schematic of an OD model. RSIOD is one of the important research directions in the field of computer vision. In RSI, due to the diversity of the scale, direction and shape of the object, as well as the complex background interference, the OD task faces many challenges3,4,5. With the development of remote sensing technology and deep learning, methods for RSIOD have emerged in an endless stream, and remarkable results have been achieved6,7,8,9,10,11,12,13,14,15,16,17. These methods effectively improve the OD performance in RSI by introducing different network structures, feature fusion strategies and optimization techniques. Zhang et al.18 can effectively detect directional objects in RSI through covariant network and equivariant contrast learning. Aiming at the problem of directional OD in RSI, this method adopts the strategy of covariant network and equivariant contrast learning, so as to improve the accuracy and accuracy of detection. Wang et al.19 introduced the improved YOLOv5n lightweight remote sensing aircraft OD network, which achieves efficient OD in RSI with low computational complexity and is suitable for resource-constrained scenarios. This method effectively improves the detection accuracy and performance of aircraft objects in RSI through the improved YOLOv5 n network. Huang et al.20 used cross-image context information to improve the accuracy and robustness of salient OD in RSI, and provided new ideas and methods for further research in the field of salient OD in RSI. Huo et al.21 improved the accuracy and performance of salient OD in RSI through global and multi-scale aggregation networks, which is suitable for RSI analysis in various complex scenes. Aiming at the problems of background interference and small OD difficulty in OD in RSI, Dong et al.22 adopted the strategy of background separation and small object compensation, which improved the accuracy and robustness of detection. By introducing explicit semantic guidance, Liu et al.23 can effectively detect small objects in RSI, improve the accuracy and efficiency of detection, and provide a new and effective method for the detection of small objects in RSI. Guo et al.24 proposed a fully deformable convolution network for ship detection in RSI. This network can better adapt to the irregular shape of the ship in the RSI, thereby improving the accuracy and robustness of the detection. By introducing deformable convolution, significant improvements have been made in the detection of objects such as ships, which provides new ideas and methods for the field of OD in RSI. Liu et al.25 used a translation window mechanism to process image information, and Swin Transformer introduced a hierarchical structure and a translation window mechanism, which effectively captured the global and local information in the image and achieved better visual representation. Ni et al.26 combined multiple feature fusion and learning methods to improve the classification accuracy of scenes in synthetic aperture radar images. It can not only improve the accuracy of image classification, but also enhance the ability to understand and characterize complex scenes in synthetic aperture radar images. It has important application value. Luo et al.27 introduced the channel attention mechanism to enhance the channel information in the Feature Pyramid Network (FPN). This mechanism can effectively improve the performance of OD, so that the model can better adapt to the OD tasks in different scales and complex scenes in RSI. Gao et al.28 proposed a method of global perception and local refinement to achieve accurate detection of objects in RSI at different scales. This method can effectively capture the global and local information in the image, and improve the detection accuracy and robustness of the object in the RSI. Ming et al.29 improved the accuracy of OD in any direction by learning dynamic anchors to adapt to objects with different angles and shapes. This strategy effectively solves the problem that the traditional fixed anchor point is difficult to adapt to various object shape and direction changes, and provides new ideas and methods for OD tasks in RSI. Qian et al.30 proposed a bounding box regression method, which aims to build a bridge between the two, to achieve information sharing and conversion between object-oriented detection and horizontal OD, and then improve the detection accuracy. Yu et al.31 introduced an algorithm of Attention-Based FPN and its application in ship detection in RSI. The A-FPN algorithm improves the performance and accuracy of ship OD in RSI through the design of feature pyramid network. These methods effectively improve the OD performance in RSI by introducing different network structures, feature fusion strategies and optimization techniques. Among them, including rotation perception and multi-scale convolution neural network32, hierarchical region generation33 and multi-scale feature fusion34, network structure with rotation perception35 and loss function, anchor-free OD method, etc.36,37,38. These methods have achieved remarkable results in different aspects, and provide effective solutions for RSIOD tasks. However, there are still some challenges, such as object scale change and complex background interference, which need to be further studied and improved.

Therefore, this paper proposes an OD algorithm called YOLO-ACPHD(YOLOv8-Adaptive Condition Awareness Technology High-Dimensional Decoupling Technology, YOLO-ACPHD). The algorithm uses ACAT to adapt to objects of different scales and locations by dynamically adjusting the parameters of the convolution kernel, solving the challenges of multi-scale objects and dense overlap of objects. Instead of using fixed convolution kernels, ACAT adaptive generates convolution kernels of different shapes and sizes by analyzing the characteristics of the input data. Using the HDDT, the calculation is performed by means of separate convolution, that is, the input data is first deeply convoluted, and then each channel is spatially convoluted. HDDT can reduce the amount of calculation to 1/N, where N is the number of convolution kernels. Through these improvements, the YOLO-ACPHD algorithm has better applicability, accuracy, and robustness in dealing with RSIOD tasks.

Data and methods

Data set introduction

Figure 2 depicts the RSOD data set, an open resource tailored for the detection of minute objects within RSI. The sample image of the RSOD RSI data set includes: aircraft, including 4993 aircraft in 446 images; oiltank, including 1586 oil tanks in 165 images; overpass, including 180 overpasses in 176 images; playground, containing 191 playgrounds in 189 images.

Figure 2
figure 2

RSOD RSI data sample diagram.

Full size image

YOLO-ACPHD module

The YOLO-ACPHD model proposed in this paper mainly expands the following techniques to improve the detection effect and robustness.

The ACAT module can dynamically adjust the parameters of the convolution kernel to adapt to objects of different scales and positions. In RSI, the scale, direction and shape of objects are different, so it is difficult to adjust them effectively based on fixed convolution function. Employing a convolution neural network that leverages the attributes of training instances, this approach facilitates the automated learning of convolution neural networks rooted in these characteristics. Furthermore, it gives rise to multi-level and multi-aspect convolution neural networks, all grounded in the nuances of the training data. it enhances the adaptability and accuracy to complex scenes.

Through the deep convolution of each channel, the HDDT model carries on the convolution operation between each channel. Compared with the conventional convolution algorithm, the computing time of HDDT algorithm is reduced to 1/N, where N is the number of convolution kernels. Optimizing computational efficacy in the handling of extensive RSI significantly accelerates object identification by minimizing computational requirements.

By using ACAT and HDDT modules, the YOLO-ACPHD RSIOD algorithm can be more flexible to adapt to objects of different scales and locations and at the same time improve the computational efficiency of the algorithm, Consequently, the accuracy of identification and the resilience are significantly improved in the context of complex and dynamic RSI.

Research on RSIOD methods based on YOLO-ACPHD variety of advanced OD techniques are used to improve the performance and robustness of the algorithm. As shown in Fig. 3.

Figure 3
figure 3

YOLO-ACPHD overall network architecture.

Full size image

ACAT and HDDT modules are used to improve the adaptability and computational efficiency of the algorithm. By adaptively adjusting the characteristics of each region and each region, the ACAT model enhances the adaptive ability and accuracy of various features. HDDT algorithm uses segmented convolution operation, which not only keeps the spatial information, but also improves the robustness to objects. The research results of this project will lay a foundation for the research of RSI data analysis and analysis methods for complex scenes.

Adaptive Condition Awareness Technology

As depicted in Fig. 4, within the standard Convolution layer, a uniform convolution kernel, referred to also as a weight, is applied across all input instances. This means that the convolution operation is the same regardless of the content of the input example. In the ACAT layer, the Convolution kernel assumes an adaptive role, transforming dynamically based on the input sample at hand. This means that each input example can have its own unique convolution kernel. The ACAT layer is designed to adaptively learn and adjust the convolution kernel according to the features and context information of the input examples, aiming to enhance the model’s capacity for expression and its generalization capabilities. This is achieved by conceptualizing the convolution kernel in relation to the input instance, the ACAT layer can share and learn different convolution kernel parameters between different examples so that the model can better adapt to the diverse input data.

Figure 4
figure 4

Conditional parameter convolution diagram.

Full size image

In ACAT, the convolution kernel is parameterized as a function of the input example. Particularly, the parameters for the convolution kernel can be derived utilizing the subsequent equation.

$$\begin{aligned} Output(x)=\sigma ((\alpha _{1} W _{1} +\cdots +\alpha _{n} \cdot W_{n} *x). \end{aligned}$$
(1)

Here x denotes the output of the previous layer, n denotes that the ACAT layer has n convolution kernels\((W_{i} )\), \(\sigma\) denotes the activation function, \(\alpha _{i}= r_{i}(x)\) represents a sample-dependent weighted parameter39.

Traditional convolution layers typically enhance their capacity through augmenting the kernel’s height/width or the quantity of Icano channels. Nevertheless, during convolution, every extra parameter necessitates more multiplication-addition operations, a computation burden that scales linearly with the image’s pixel count. In the ACAT layer, before applying convolution, a convolution kernel is calculated for each example as a linear combination. Each convolution kernel only needs to be calculated once, but it is applied to many different positions of the input image. The integration of n enhances the overall capability of the network, and this augmentation demands a mere insignificant computational expense. Each additional parameter requires an additional multiplication operation. As you can see from Fig. 5, the ACAT layer is mathematically equivalent to a high-cost expert linear mixed equation in which each expert has a static convolution:

Figure 5
figure 5

Expert linear mixed diagram.

Full size image
$$\begin{aligned} \sigma ((\alpha _{1}\cdot W_{1}+\cdots +\alpha _{n}\cdot W_{n}) *x)=\sigma (\alpha _{1}\cdot (W_{1}*x)+\cdots +\alpha _{n}\cdot (W_{n}*x)). \end{aligned}$$
(2)

Therefore, ACAT has the same capacity as the linear mixed expert formula of n experts, but it is computationally efficient because it only needs to calculate one expensive convolution. This formula delves into the nature of ACAT and links it to previous work on conditional calculations and expert mixing. In ACAT, the path selection function of each instance is very critical.

A routing function is designed that has high computational efficiency, can effectively distinguish input examples, and is easy to explain. The algorithm uses three steps to solve the problem, namely: global average common, fully connected layer, Sigmoid activation. The route weighting is calculated as follows:

$$\begin{aligned} r(x)=Sigmoid(GlobalAveragePool(x)R). \end{aligned}$$
(3)

For the input x, the global average pooling is first performed, and then the right is multiplied by a matrix R (the purpose of the matrix is to map the dimensions to n experts to achieve the subsequent linear combination). Finally, the weights on each dimension are reduced to the [0, 1] interval by sigmoid. Therefore, different routing weight vectors are obtained according to the different input x.

High-dimensional decoupling technology

Based on the HDDT, the standard \(3 \times 3\) convolution operation is decomposed into two steps: deep convolution and point convolution, as shown in Fig. 6. The traditional \(3 \times 3\) convolution performs both channel and spatial direction calculations in one step. The high-dimensional decoupling technique decomposes the calculation into two steps: deep convolution and point convolution.

Each input channel undergoes an individual convolution process with its exclusive convolution filter. Consequently, every channel is equipped with its own convolution kernel for distinct data processing, thereby conducting the convolution operation across the channel dimension. Point convolution is used to create a linear combination of deep convolution outputs. Point convolution uses \(1 \times 1\) convolution kernels to linearly combine the output channels of deep convolution. This process can be seen as the dimension reduction and fusion of channel dimensions.

Standard convolution. Assuming that the convolution kernel size is \(D_{k}\times D_{k}\) , the input channel is M, the output channel is N, and the output feature map size is \(D_{F}\times D_{F}\), then after the standard convolution, it can be calculated, the number of parameters is \(D_{k}\times D_{k} \times M \times N\), the amount of calculation is \(D_{k}\times D_{k} \times M \times N \times D_{F}\times D_{F}\).

Deep convolution. The convolution kernel of deep convolution is a single channel mode, and each channel of the input needs to be convoluted, so that the output feature map with the same number of channels as the input feature map will be obtained. That is, the number of input feature map channels = the number of convolution kernels = the number of output feature maps. The convolution size of deep convolution is \(D_{k}\times D_{k} \times 1\), the number of convolution kernels is M, and each must do \(D_{F}\times D_{F}\) multiplication operations. Parameter is \(D_{F}\times D_{F} \times M\), calculation amount is \(D_{k}\times D_{k} \times M \times D_{F}\times D_{F}\).

Pointwise convolution. Pointwise convolution W/H dimension unchanged, change the channel. According to deep convolution, the number of input feature map channels = the number of convolution kernels = the number of output feature maps, which will lead to too few output feature maps (too few channels of output feature maps, which can be regarded as the number of output feature maps is 1 and the number of channels is 3), which may affect the effectiveness of information. At this time, point-by-point convolution is needed. Point wise Convolution (PWConv) is essentially a \(1\times 1\) convolution kernel for dimension elevation. The convolution size of point-by-point convolution is \(1\times 1 \times M\) the number of convolution kernels is N, and each must do \(D_{F}\times D_{F}\) multiplication operations. Parameter is \(M\times N\), calculation amount is \(M \times N \times D_{F}\times D_{F}\).

HDDT consists of deep convolution and point-by-point convolution. Deep convolution is used to extract spatial features, and point-by-point convolution is used to extract channel features. HDDT groups convolutions on the feature dimension, performs independent deep convolutions on each channel, and aggregates all channels using a \(1\times 1\) convolution before output. Parameter is \(D_{k}\times D_{k} \times M + M \times N\), calculation amount is \(D_{k}\times D_{k} \times M \times D_{F}\times D_{F} + M \times N \times D_{F}\times D_{F}\).

Comparison between standard convolution and HDDT.

$$\begin{aligned} Parameter quantity ratio= & \frac{HDDT}{Conv} = \frac{D_{k}\times D_{k} \times M \times N }{D_{k}\times D_{k} \times M + M\times N}=\frac{1}{N}+\frac{1}{{D_{k}^{2} } }. \end{aligned}$$
(4)
$$\begin{aligned} Calculation ratio= & \frac{HDDT}{Conv} = \frac{D_{k}\times D_{k} \times M \times N \times D_{F}\times D_{F} }{D_{k}\times D_{k} \times M \times D_{F}\times D_{F} + M\times N \times D_{F}\times D_{F}}=\frac{1}{N}+\frac{1}{{D_{k}^{2}} }. \end{aligned}$$
(5)

In general, N is larger, \(\frac{1}{N}\) is negligible, \(D_{k}\) represents the size of the convolution kernel, if \(D_{k}\), \(\frac{1}{{D_{k}^{2}}}=\frac{1}{9}\), if we use the convolution kernel of the commonly used \(3\times 3\), then the number of parameters and calculations using HDDT is reduced to about one-ninth of the original.

In summary, the deep separable convolution based on high-dimensional decoupling technology decomposes the standard \(3\times 3\) convolution into two steps: deep convolution and point convolution. This decomposition can improve computational efficiency and make better use of model parameters and computing resources40.

Figure 6
figure 6

Expert linear mixed diagram.

Full size image

When the standard convolution filter K of size \(W^{\prime }\times H^{\prime } \times M\times N\) is applied to the input feature mapping F of size \(W\times H \times M\), an output feature mapping O of size \(D_{f}\times D_{f}\times M\) can be obtained. In standard convolution, each filter K is a tensor of size \(W^{\prime }\times H^{\prime } \times M\). The filter is convoluted with the input feature mapping F to obtain an output feature mapping O with a size of \(D_{f}\times D_{f}\times M\), where N is the number of filters. Each slice of the filter K with a size of \(D_{f}\times D_{f}\times N\) is multiplied by the input feature map F element by element, and the results are added to obtain a feature map with a size of \(D_{f}\times D_{f}\times M\).

In summary, a standard convolution filter of size \(W^{\prime }\times H^{\prime } \times M\times N\) is applied to the input feature mapping of size \(W\times H \times M\), and an output feature mapping of size W can be generated. In the context of Convolution technique, individual units multiply the input feature map, cumulatively aggregating the values across each section to derive the resulting output.

$$\begin{aligned} O_{k,l,n}=\sum _{i,j,m} K_{i,j,m,n}\cdot F_{k+i-1,l+j-1,m}. \end{aligned}$$
(6)

In the high-dimensional decoupling technology, the above calculation is decomposed into two steps. The first step is to apply \(3\times 3\) deep convolution K to each input channel. Specifically, for the input feature mapping F with a size of \(W\times H \times M\), \(3\times 3\) deep convolution kernels are used to perform convolution operations with each channel of the input feature mapping. In this way, M output feature maps of size \(W\times H\) will be obtained.

$$\begin{aligned} \hat{O} _{k,l,n}=\sum _{i,j,m} \hat{K} _{i,j,m}\cdot F_{k+i-1,l+j-1,m}. \end{aligned}$$
(7)

The second step is point convolution, which is used to create a linear combination of deep convolution outputs. Using the convolution kernel of \(1\times 1\), the M output feature maps are linearly combined by point-by-point convolution \(\hat{K}\) to obtain an output feature map of \(W\times H\) size. This process is equivalent to dimensionality reduction and fusion of channel dimensions.

$$\begin{aligned} O_{k,l,n}=\sum _{i,j,m} \hat{K} _{m,n}\cdot \hat{O}_{k-1,l-1,m}. \end{aligned}$$
(8)

Deep convolution and point convolution play different roles in generating new features. In this context, deep convolution serves to discern the spatial relationships inherent within an image, essentially extracting details such as edges, textures, and other pertinent information. By employing profound convolution across channels, the algorithm enhances its capacity to discern the image’s spatial architecture and distill a wealth of information from it. Conversely, point convolution is typically employed to seize the interdependencies that exist among distinct channels. By using the convolution kernel function of \(1\times 1\) and the point convolution method, the channel dimensions of each characteristic graph are linearly synthesized to achieve the interaction and fusion between channels. This operation can help the network learn the correlation and importance between different channels, thereby generating new features with more representational capabilities.

Results

Index curve analysis

Analysis of F1 value curve

As shown in Fig. 7, the F1 value curve comparison of YOLOv8 series algorithms in RSIOD task is shown. F1 value is the comprehensive evaluation index of Precision and Recall, which is the harmonic mean of the two. The F1 value comprehensively considers the accuracy and recall of the classifier, and gives the same weight to Precision and Recall. The computational expression for the F1 value is as follows.

$$\begin{aligned} F1=2\times (precision\times Recall)/(precision+Recall). \end{aligned}$$
(9)

The value of F1 is between 0 and 1, and the closer the value is to 1, the better the performance of the model is. The F1 value is suitable for the case of an unbalanced class distribution. When the performance of the classifier is quite different in positive and negative examples, the F1 value can better evaluate the performance of the model.

Figure 7
figure 7

F1 value curve.

Full size image

Among them, Fig. 7a YOLOv8 algorithm has the best F1 value of 0.93 when the Confidence value is 0.502; Fig. 7b YOLOv8MobileOne algorithm has the best F1 value of 0.99 when the Confidence value is 0.518; Fig. 7c YOLOv8SWinTransformer algorithm has the best F1 value of 0.92 when the Confidence value is 0.471; Fig. 7d YOLOv8-GCE algorithm has the best F1 value of 0.99 when the Confidence value is 0.523; Fig. 7e YOLOv8-GCOD algorithm has the best F1 value of 0.98 when the Confidence value is 0.627; Fig. 7f YOLO-ACPHD algorithm has the best F1 value of 0.99 when the Confidence value is 0.531.

Precision value curve analysis

As shown in Fig. 8, the Precision value curve comparison of the YOLOv8 series algorithms in the RSIOD task is shown. Precision: Precision refers to the proportion of positive samples predicted by the classifier to the actual positive samples. It measures the accuracy of the model in samples that are predicted to be positive. The calculation formula is as follows.

$$\begin{aligned} Precision=TP/(TP+FP). \end{aligned}$$
(10)
Figure 8
figure 8

Precision value curve.

Full size image

Among them, Fig. 8a YOLOv8 algorithm has the best Precision value of 0.99 when the Confidence value is 1; Fig. 8b YOLOv8MobileOne algorithm has the best Precision value of 1 when the Confidence value is 0.917; Fig. 8c YOLOv8SWinTransformer algorithm has the best Precision value of 0.98 when the Confidence value is 1; Fig. 8d YOLOv8-GCE algorithm has the best Precision value of 1 when the Confidence value is 0.911; Fig. 8e YOLOv8-GCOD algorithm has the best Precision value of 1 when the Confidence value is 0.924; Fig. 8f. The YOLO-ACPHD algorithm has the best Precision value of 1 when the Confidence value is 0.947.

Precision-Recall value curve analysis

As shown in Fig. 9, the Precision-Recall value curve comparison of the YOLOv8 series algorithms in the RSIOD task is demonstrated. Precision-Recall curve. Precision-Recall curve is an evaluation index that comprehensively considers Precision and Recall. It draws a curve with Recall as the abscissa and Precision as the ordinate. The coordinates of each point on the curve represent the accuracy under different recall rates. The purpose of plotting such a curve is to visualize how the model’s effectiveness fluctuates across varying classification thresholds. Ideally, an optimal model performance is indicated when the curve hugs the top-right corner, implying high Precision and Recall values.

Figure 9
figure 9

Precision-Recall value curve.

Full size image

Among them, Fig. 9a YOLOv8 algorithm has the best Precision-Recall value of 0.994 at mAP@0.5; Fig. 9b YOLOv8MobileOne algorithm has the best Precision-Recall value of 0.992 at mAP@0.5; Fig. 9c YOLOv8SWinTransformer algorithm has the best Precision-Recall value of 0.945 at mAP@0.5; Fig. 9d YOLOv8-GCE algorithm has the best Precision-Recall value of 0.993 at mAP@0.5; Fig. 9e YOLOv8-GCOD algorithm has the best Precision-Recall value of 0.994 at mAP@0.5; Fig. 9f YOLO-ACPHD algorithm has the best Precision-Recall value of 0.994 at mAP@0.5.

Recall value curve analysis

As shown in Fig. 10, the Recall value curve comparison of the YOLOv8 series algorithms in the RSIOD task is shown. Recall: Recall refers to the proportion of samples that are correctly predicted as positive by the classifier. It measures the recall rate of the model for positive samples. The mathematical expression for its computation is as follows.

$$\begin{aligned} Recall=TP/(TP+FN). \end{aligned}$$
(11)
Figure 10
figure 10

Recall value curve.

Full size image

Among them, Fig. 10a YOLOv8 algorithm has the best Recall value of 0.98 when the Confidence value is 0.0; Fig. 10b YOLOv8MobileOne algorithm has the best Recall value of 1 when Confidence value is 0.0; Fig. 10c YOLOv8SWinTransformer algorithm has the best Recall value of 0.98 when the Confidence value is 0.0; Fig. 10d YOLOv8-GCE algorithm has the best Recall value of 1 when the Confidence value is 0.0; Fig. 10e YOLOv8-GCOD algorithm has the best Recall value of 1 when the Confidence value is 0.0; Fig. 10f YOLO-ACPHD algorithm has the best Recall value of 1 when the Confidence value is 0.0.

Comparative experiments

As shown in Fig. 11, the comparison results of the mean Average Precision (mAP) indicators of different OD models on the RSOD RSI data set. This dot plot clearly reflects the performance differences of various models on the RSOD data set. The YOLO series models (YOLOv3, YOLOv4, YOLOv5s, YOLOv7 and YOLOv8) performed well in the RSOD task. In particular, two modules (ACAT and HDDT) are added to the YOLOv5s and YOLOv7 models respectively, which significantly improves the performance of the original model and achieves a higher mAP value. Among them, the mAP of YOLOv7-HDDT model is 98.56\(\%\), which shows its strong ability in the field of RSIOD. In addition, YOLO-ACPHD(ours) performed particularly well in the RSOD task, with a mAP value of 99.36\(\%\). This shows that our model has achieved significant advantages. In addition to the YOLO series models, other models (CBD-E, CornerNet, MFC and RetinaNet) also show certain performance. However, compared with the YOLO series models, their mAP values are generally low, indicating that the YOLO series models have better generalization ability and robustness in RSIOD tasks.

Figure 11
figure 11

Contrast experimental point-line diagram of RSOD RSI.

Full size image
Table 1 Comparative experimental results of RSOD RSI.
Full size table

Table 1 shows in detail the comparative experimental results of different YOLO series and other OD algorithms on the RSOD RSI data set. These algorithms are evaluated on four key metrics. mAP/\(\%\), Parameters/M, FLOPs/G and FPS. It can be seen from the table that different algorithms show significant differences in performance. In terms of mAP, YOLOv5s, various improvements of YOLOv5s (ACAT, HDDT, SA, CSSGD), YOLOv7 and its improvements (ACAT, HDDT), and YOLO-ACPHD(ours) all showed high accuracy, exceeding 95\(\%\), of which YOLO-ACPHD(ours) reached the highest 99.36\(\%\). This shows that the proposed algorithm has good recognition ability in the OD task of RSI. In terms of Parameters/M and FLOPs/G, although high precision is usually accompanied by high computational costs, some algorithms such as YOLOv8, YOLOv5s, YOLOv7 and YOLO-ACPHD(ours) also control the increase in the number of parameters and floating-point operations while maintaining high precision, showing better computational efficiency. In particular, YOLO-ACPHD(ours) has a relatively low number of parameters and floating-point operations while maintaining the highest accuracy. In terms of FPS, this is an important indicator to measure the real-time performance of the algorithm. It can be seen from the table that YOLOv8, YOLOv5s, YOLOv7 and YOLO-ACPHD(ours) all have high FPS values, exceeding 100 frames/second, which indicates that these algorithms have good real-time performance in OD tasks of RSI. In particular, YOLO-ACPHD(ours) achieves the highest FPS value of 141.38 frames/s while maintaining high accuracy and low computational cost. It provides an efficient and accurate solution for the OD task of RSI.

Ablation experiment

As shown in Fig. 12, we conducted detailed ablation experiments to evaluate the impact of different modules on the performance of the YOLOv8 model on the RSOD RSI dataset. The experimental results show that the performance of the YOLOv8 model has been significantly improved after the introduction of HDDT, from the original 95.06\(\%\) mAP to 99.20\(\%\). This improved version is named YOLOv8-HDDT. Further, we incorporated ACAT into YOLOv8-HDDT to obtain YOLO-ACPHD(ours). As shown in the dot-line diagram in Fig. 12, YOLO-ACPHD(ours) reaches 99.36\(\%\) on the mAP performance index, which once again proves the effectiveness of the ACAT module in improving the detection accuracy of the model. Figure 12 intuitively shows the change trend of YOLOv8 ’s mAP performance index on the RSOD RSI data set when gradually adding different modules. Through this series of experiments, we have fully verified the significant role of HDDT and ACAT in improving the accuracy of remote sensing OD.

Figure 12
figure 12

Ablation experimental point-line diagram of RSOD RSI.

Full size image
Table 2 The ablation experimental results of RSOD RSI.
Full size table

Table 2 shows in detail the performance comparison results of different modules based on the YOLOv8 basic model on the RSOD RSI data set. It can be seen from the table that with the gradual optimization of the YOLOv8 model, the performance of the model has been improved on multiple key indicators. In terms of mAP/\(\%\), YOLOv8-HDDT has a significant improvement compared with the original YOLOv8 model, from 95.06 to 99.20\(\%\), showing the positive impact of HDDT module on the accuracy of model detection. The final YOLO-ACPHD(ours) further improved the mAP to 99.36\(\%\), indicating that the introduction of the ACAT module brought more accurate detection results to the model. In terms of Parameters/M, although the performance of the model has been improved, both YOLOv8-HDDT and YOLO-ACPHD(ours) have successfully maintained relatively low parameters, which are 5.40 M and 5.42 M, respectively, compared with the original YOLOv8 (5.92M). This shows that the model maintains good computational efficiency while improving performance, which is conducive to deployment in practical applications. In terms of FLOPs/G, YOLOv8-HDDT and YOLO-ACPHD(ours) also showed optimization effects, and FLOPs decreased from the original 13.77 G to 12.69 G and 12.47 G, respectively. This means that the amount of calculation required for the model to perform forward propagation is reduced, and the running speed of the model is further improved. In terms of FPS, YOLOv8-HDDT and YOLO-ACPHD(ours) have achieved great improvement. YOLOv8-HDDT improves FPS from 126.50 to 137.21, and YOLO-ACPHD(ours) reaches 141.38, indicating that the model achieves faster detection speed while maintaining high accuracy and low computational cost, and meets the needs of real-time detection of RSI.

Experimental results

Figure 13
figure 13

RSOD RSI detection effect diagram.

Full size image

As shown in Fig. 13, the prediction results of the YOLO-ACPHD algorithm model designed in this paper on the RSOD data set show remarkable results. In this image, the (a) section clearly shows the label sketch map of each object in the data set, providing us with a benchmark to compare and evaluate the performance of the algorithm. The (b) part intuitively shows the actual detection effect of the algorithm after running on the verification set. In order to fully demonstrate the accuracy and reliability of the algorithm, we specially selected the most challenging object-the aircraft object which is the most difficult to detect completely as an example. By showing the comparison of 12 sets of labels and detection effects, we can clearly see that the YOLO-ACPHD algorithm proposed in this paper shows extremely high accuracy in both the location of the object and the shape recognition of the object. Almost every aircraft object is accurately detected by the algorithm without deviation. This achievement not only proves the powerful performance of the algorithm on known data sets, but more importantly, it shows the excellent generalization ability of the algorithm on unseen data. Whether in the training set or in new and unknown images, the algorithm can detect and locate the object object stably and accurately, thus playing a huge role in practical applications.

Discussion

In this paper, the YOLO-ACPHD algorithm model is proposed. By using adaptive conditional sensing technology and HDDT, the challenges of multi-scale objects, dense overlap of objects and uneven data distribution in RSIOD are significantly improved. The experimental results on the RSOD data set show that the YOLO-ACPHD model has an F1 value of 0.99, a Precision value of 1, a Precision-Recall value of 0.994, a Recall value of 1, and a mAP value of 99.36\(\%\), showing the highest performance. It is considered to be an effective method and can achieve better performance on the RSOD RSI data set. However, future research can further optimize the adaptive conditional sensing technology to improve the detection effect of small objects, and improve the operation of HDDT to reduce the amount of calculation and improve the speed. At the same time, network pruning, quantization and other methods can be introduced to reduce the complexity of the model, and other OD techniques and model structures can be explored to improve the accuracy and robustness of the algorithm.