False target suppression method based on multi-channel detection association

Abstract

The invention discloses a false target suppression method based on multi-channel detection association, which comprises the following steps of: firstly completing respective channel image target detection and identification on an infrared channel and a visible light channel, normalizing image coordinates of all detection results, and then converting the target detection and identification results of each channel into a unified image association space; in the image correlation space, maximum value equalization operation is carried out on all detection results, false target suppression is achieved, and according to the mapping relation of each channel image in the image correlation space, the final detection result is converted and then output to each channel. According to the method, the generation of false targets can be remarkably inhibited, and the problem that the detection and recognition performance of photoelectric equipment is reduced in a complex scene can be effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for suppressing the generation of false targets when a photoelectric device has a target detection and recognition function. Background Art

[0002] Currently, with the support of large-scale data sets, intelligent target detection and recognition technology has been widely used in optoelectronic equipment such as security, traffic monitoring, and detection, and has achieved quite significant practical application effects.

[0003] Intelligent target detection and recognition technology usually adopts an end-to-end neural network model structure, performing a series of convolution, pooling, fusion and downsampling operations on the input image data to obtain neural network feature maps at multiple scales. Then, linear regression is performed on the deep abstract feature maps to obtain information such as the presence, category and location of the target. Finally, the final detection result is generated through post-processing such as non-maximum suppression.

[0004] In practical device applications, intelligent object detection and recognition technology often suffers from insufficient generalization performance because model training cannot guarantee full coverage of all practical application scenarios. For example, when identifying objects in complex environments, complex background features can easily generate false targets. This creates false target detection results within the background area, reducing the device's detection and recognition performance in complex scenarios. Summary of the Invention

[0005] The technical problem to be solved by the present invention is that the model in complex scenarios is prone to generate false targets, resulting in a decrease in the detection and recognition performance of the device.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a false target suppression method based on multi-channel detection association, which is used in optoelectronic equipment with infrared and visible light two-channel imaging, comprising the following steps:

[0007] S1 deploys an intelligent detection and recognition model to perform target detection and recognition on the infrared channel input image, and obtains the infrared channel detection results by sorting them from high to low according to the confidence level conf:

[0008]

[0009] Deploy an intelligent detection and recognition model to perform target detection and recognition on images input through the visible light channel. Sorting the confidence levels conf from high to low yields the visible light channel detection results:

[0010]

[0011] S2, normalize the image coordinates of all infrared channel and visible light channel detection and recognition results: According to the resolution W×H of each channel image, the position coordinates of the detection results are normalized to obtain the infrared channel detection results respectively:

[0012]

[0013] Visible light channel detection results:

[0014]

[0015] S3, the horizontal and vertical field angles of the infrared channel image are recorded as (ω ir , γ is ), the horizontal and vertical field angles of the visible light channel image are recorded as (ω vir , γ vis ), the image association space field of view angle range is the maximum field of view angle range of all sensors, and the target detection and recognition results of each channel are converted to a unified image association space by the following formula: ω=Max(ω ir ,ω vis ), γ=Max(γ ir , γ vis );

[0016] S4, performs maximum equalization operation on all detection results in the image association space to suppress false targets;

[0017] S5, according to the mapping relationship between the infrared channel image and the visible light channel image in the image association space, respectively, through the formula And the formula The final detection results are output to the infrared channel and the visible light channel, and false target suppression is completed in the output results of each channel.

[0018] Furthermore, the infrared channel detection result in step S3 is transformed into

[0019]

[0020] Furthermore, the visible light detection result in step S3 is transformed into

[0021]

[0022] Furthermore, the step S4 is specifically as follows:

[0023] S41, taking the first detection result of the infrared channel in the image correlation space, and comparing it with all the detection results of the visible light channel in sequence to determine whether there is an overlapping area between the two detection results;

[0024] S42, by formula Det mean =(Det ir +Det vis ) / 2 Calculate the maximum value equalization detection result of the first overlapping area of ​​the visible light channel and the current infrared channel detection result, and mark the detection result of the visible light channel that has undergone maximum value equalization;

[0025] S43, continue to compare with subsequent detection results of the visible light channel. When there is an overlapping area between the two detection results, delete the corresponding overlapping detection result in the visible light channel detection result sequence until all detection results are compared;

[0026] S44, after comparing all the detection results of the visible light channel, if there is no detection result of the overlapping area, the detection result of the maximum equalization of the current infrared detection result is Det mean =(De tir ) / 2;

[0027] S46, taking the next infrared detection result of the infrared channel in the image correlation space, and repeating steps S41 to S44 to complete the maximum value equalization of all infrared channel detection results;

[0028] S47, by formula Det mean =(Det vis ) / 2 equalizes the results of the visible light channel that have not been equalized to the maximum value. At this point, the maximum value equalization of all results in the image correlation space is completed.

[0029] The beneficial effects of the present invention are as follows: the false target suppression method of the present invention is used to solve the problem that optoelectronic equipment is prone to generating false targets. Compared with directly using an end-to-end neural network to detect and identify single-channel image targets, it utilizes the detection results of multiple channels and performs maximum equalization in a unified correlation space, which can significantly suppress the generation of false targets and effectively solve the problem of decreased detection and recognition performance of optoelectronic equipment in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a processing flow chart of the method of the present invention;

[0031] Figure 2 It is a schematic diagram of overlapping discrimination of detection results. DETAILED DESCRIPTION

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0033] The false target suppression method based on multi-channel detection association is applied to actual optoelectronic devices that include multi-channel imaging. Typical multi-channel optoelectronic devices include at least two imaging sensors, infrared and visible light. Therefore, n is generally ≥ 2. Therefore, this article describes the specific implementation method using an optoelectronic device with two channels, infrared and visible light.

[0034] Reference Figure 1 As shown, this embodiment discloses a false target suppression method based on multi-channel detection association, and the steps are as follows.

[0035] (1) The optoelectronic equipment deploys intelligent detection and recognition models for the infrared channel and the visible light channel respectively, and independently completes the detection and recognition of the image targets in each channel.

[0036] All detection results of each channel are sorted from high to low according to the confidence level conf, so the infrared channel detection results are recorded as follows.

[0037]

[0038] The visible light channel detection results are recorded as follows.

[0039]

[0040] (2) The image coordinates of all detection results are normalized.

[0041] According to the resolution W×H of each channel image, the position coordinates of the detection results are normalized. The infrared channel detection results and the visible light channel detection results are as follows.

[0042] The infrared channel detection results are recorded as follows.

[0043]

[0044] The visible light channel detection results are recorded as follows.

[0045]

[0046] (3) Convert the target detection and recognition results of each channel into a unified image association space.

[0047] The horizontal and vertical field angles of the infrared channel image are recorded as (ω ir , γ ir ), the horizontal and vertical field angles of other channel images such as the visible light channel image are recorded as (ω vis , γ vis ).

[0048] The field of view angle range of the image correlation space is the maximum field of view angle range of all sensors, which is calculated as follows:

[0049] ω=Max(ω ir ,ω vis );γ=Max(γ ir , γ vis ).

[0050] Therefore, the calculation results are as follows.

[0051] The infrared channel detection results are recorded in the image association space as follows.

[0052]

[0053] The visible light channel detection results are recorded in the image association space as follows.

[0054]

[0055]

[0056] At this point, all channel detection results are converted to the same image association space.

[0057] (4) In the image association space, all detection results are subjected to maximum equalization to suppress false targets. The specific steps are as follows.

[0058] (41) Take the first detection result of the infrared channel in the image association space.

[0059] (42) Compare with all the detection results of the visible light channel in turn to see if there is any overlapping area between the two detection results, such as Figure 2 shown.

[0060] (43) The first overlapping detection result of the visible light channel and the current infrared detection result are calculated to obtain the maximum value equalization detection result, as follows: Det mean =(Det ir +Det vis ) / 2; and at the same time mark the detection results of the visible light channel that has undergone maximum value equalization.

[0061] (44) Continue to compare with the subsequent detection results of the visible light channel. When the two detection results overlap again, delete the corresponding overlapping detection results in the visible light channel detection result sequence until all detection results are compared.

[0062] (45) After comparing all the detection results of the visible light channel, if there is no overlapping detection result, the detection result of the maximum equalization of the current infrared detection result is: Det mean =(De tir ) / 2.

[0063] (46) Take the next infrared detection result of the infrared channel in the image association space, repeat steps (42) to (45), and complete the maximum value equalization of all infrared channel detection results.

[0064] (47) The results of visible light channel without maximum equalization are equalized as follows:

[0065] Det mean =(Det vis ) / 2

[0066] At this point, the maximum equalization of all results in the image correlation space is completed.

[0067] (5) According to the mapping relationship between the images of each channel in the image association space, the final detection result is converted and output to each channel.

[0068] The infrared channel conversion formula is:

[0069]

[0070] The visible light channel conversion formula is:

[0071]

[0072] False target suppression is accomplished in each channel output result.

[0073] The above embodiments are merely illustrative of the principles and effects of the present invention, as well as some embodiments of its application. A person skilled in the art may make several modifications and improvements without departing from the inventive concept of the present invention, and all of these modifications and improvements fall within the scope of protection of the present invention.

Claims

1. A false target suppression method based on multi-channel detection association, used for optoelectronic equipment with infrared and visible light two-channel imaging, characterized in that: The following steps are included S1, perform target detection and recognition on the image input by the infrared channel, and sort the infrared channel detection results from high to low according to the confidence level conf: Perform target detection and recognition on the image input by the visible light channel, and sort the visible light channel detection results from high to low according to the confidence level conf: S2, according to the resolution W×H of each channel image, the position coordinates of the detection results are normalized to obtain the infrared channel detection results: Visible light channel detection results: S3, the horizontal and vertical field angles of the infrared channel image are recorded as (ω ir , γ ir ), the horizontal and vertical field angles of the visible light channel image are recorded as (ω vis , γ vis ), the field of view angle range of the image association space is the maximum field of view angle range of all sensors. The target detection and recognition results of each channel are converted to a unified image association space through the following formula: ω=Max(ω ir ,ω vis ), γ=Max(γ ir , γ vis ); S4, performs maximum equalization operation on all detection results in the image association space to suppress false targets; S5, according to the mapping relationship between the infrared channel image and the visible light channel image in the image association space, respectively, through the formula And the formula The final detection results are output to the infrared channel and the visible light channel, and false target suppression is completed in the output results of each channel.

2. According to the method of claim 1, the false target suppression method based on multi-channel detection association is characterized in that: In step S3, the infrared channel detection result is transformed into 3. The false target suppression method based on multi-channel detection association according to claim 1 is characterized in that: The visible light detection result in step S3 is transformed into 4. A false target suppression method based on multi-channel detection association according to claim 1, 2 or 3, characterized in that: The step S4 is specifically as follows: S41, taking the first detection result of the infrared channel in the image association space, and comparing it with all the detection results of the visible light channel in sequence to determine whether there is an overlapping area; S42, by formula Det mean =(det ir +Det vis ) / 2 Calculate the maximum value equalization detection result of the first overlapped area of ​​the visible light channel and the current infrared channel detection result, and mark the detection result of the visible light channel that has undergone maximum value equalization; S43, continue to compare with the subsequent detection results of the visible light channel, and when there is an overlapping area between the two detection results, delete the corresponding overlapping detection result in the visible light channel detection result sequence until all detection results are compared; S44, after comparing all the detection results of the visible light channel, if there is no detection result of the overlapping area, the detection result of the maximum value equalization of the current infrared detection result is Det mean =(Det ir ) / 2; S46, taking the next infrared detection result of the infrared channel in the image association space, repeating steps S41 to S44, and completing the maximum value equalization of all infrared channel detection results; S47, by formula Det mean =(Det vis ) / 2 equalizes the results of the visible light channel that have not been equalized to the maximum value, and completes the maximum value equalization of all results in the image correlation space.

Images