CN-122023877-A - Method, device, equipment and medium for detecting target feature fusion enhancement of monitoring area

CN122023877ACN 122023877 ACN122023877 ACN 122023877ACN-122023877-A

Abstract

The invention discloses a method, a device, equipment and a medium for detecting target feature fusion enhancement of a monitoring area, which comprise the steps of acquiring images of different time sequences of the monitoring area, inputting the images into a feature pyramid network for feature coding so as to extract features of different scales and perform feature fusion; the method comprises the steps of obtaining center point coordinates of a target area, of which the area is smaller than a preset area threshold, of a label image, carrying out two-dimensional Gaussian fitting on the center point area according to the center point coordinates of the target area to generate small target area characteristic information, carrying out layered fusion on the small target area characteristic information and characteristic information extracted by different scales respectively to fuse the small target area characteristic information to corresponding decoding characteristic layers, carrying out layered decoding on each decoding characteristic layer and carrying out layered supervision on each decoding characteristic layer to judge whether a target in the target area is newly increased or not and the change state of the target area according to pixel classification. The invention can enhance the characteristic information of the small area and improve the detection precision and the anti-interference performance of the small target.

Inventors

SHENG YAMING
FAN SHAOSHENG
OUYANG FENG
HUANG MINGXING
SUN WENMIN

Assignees

中电昱创(苏州)智能科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. The method for detecting the fusion enhancement of the target characteristics of the monitoring area is characterized by comprising the following steps: acquiring images of different time sequences of a monitoring area, inputting the images into a feature pyramid network to perform feature coding so as to extract features of different scales and perform feature fusion; Inputting a label image, and acquiring the center point coordinates of a target area, the area of which is smaller than a preset area threshold, in the label image; performing two-dimensional Gaussian fitting on the central point region according to the central point coordinates of the target region to generate small target region characteristic information; carrying out layered fusion on the characteristic information of the small target area and the characteristic information extracted by different scales respectively so as to fuse the characteristic information of the small target area to a corresponding decoding characteristic layer; and carrying out layered decoding and layered supervision on each decoding characteristic layer so as to judge whether the target in the target area of the monitoring area is newly increased or not and the change state of the target area according to pixel-by-pixel classification.
2. The method for detecting the fusion enhancement of the target characteristics of the monitored area according to claim 1, wherein two-dimensional gaussian fitting is performed on the coordinates of the central point of the target area by using a two-dimensional gaussian function, and data in a specified range near the coordinates of the central point of the target area in the data obtained by fitting are not 0, and data in other areas are 0.
3. The method for detecting the fusion enhancement of the target characteristics of the monitored area according to claim 2, wherein the two-dimensional gaussian fitting is performed on the coordinates of the central point of the target area according to the following formula: Wherein, the Representing small target area characteristic information generated by fitting, , The pixel point position coordinates of the image are respectively, 、 For the preset coefficients of the two-dimensional gaussian function, , The coordinates of the center point of the target area are respectively obtained.
4. The method for detecting the fusion enhancement of the target features of the monitored area according to claim 2, wherein when a plurality of target areas with areas smaller than a preset area threshold exist, the small target area feature information generated by fitting each target area is combined to form final small target area feature information.
5. The method for detecting the target feature fusion enhancement of the monitoring area according to claim 1, wherein the step of carrying out layered fusion on the small target area feature information and the feature information extracted in different scales comprises the steps of carrying out fusion on the small target area feature information and the feature layer difference values extracted in different scales to form fusion features, and merging the fusion features into the fusion feature layers of each layer to form the fusion feature layer.
6. The method for detecting target feature fusion enhancement of a monitored area according to any one of claims 1 to 5, wherein the performing hierarchical decoding on each of the decoded feature layers includes: obtaining a decoding characteristic layer, convolving and sigmoid activating, and then inverting an output result; And taking a fusion result obtained after the small target region characteristic information is fused with a sigmoid activation layer result as a current layer mask, and taking the fusion result fused with a current layer input as the input of a next decoding layer.
7. The method for detecting target feature fusion enhancement in a monitored area according to any one of claims 1 to 5, wherein performing hierarchical supervision on each of the decoded feature layers includes: Calculating the loss function of each decoding layer by using the labels, and weighting the loss function of each layer to obtain a weighted loss function; Performing two classification on the weighting loss function pixel by pixel, and judging whether each pixel point is changed or not; and judging whether the target in the target area of the monitoring area is newly increased or not and whether the area is changed or not according to the judging result of each pixel point.
8. The utility model provides a monitoring area target feature fuses reinforcing detection device which characterized in that includes: The feature coding module is used for acquiring images of different time sequences of the monitoring area, inputting the images into the feature pyramid network for feature coding so as to extract features of different scales and perform feature fusion; The central point coordinate acquisition module is used for acquiring the central point coordinate of a target area with the area smaller than a preset area threshold value in the tag image by inputting the tag image; the data fitting module is used for carrying out two-dimensional Gaussian fitting on the central point area according to the central point coordinates of the target area to generate small target area characteristic information; The layering fusion module is used for carrying out layering fusion on the characteristic information of the small target area and the characteristic information extracted by different scales respectively to obtain enhanced characteristic information; And the layered decoding module is used for carrying out layered decoding on the enhanced characteristic information, and judging whether the target in the target area of the monitoring area is newly increased or not and the change state of the target area by pixel classification.
9. An electronic device comprising a processor and a memory for storing a computer program, wherein the processor is configured to execute the computer program to perform the method for detecting fusion enhancement of target features of a surveillance area according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the monitoring area target feature fusion enhancement detection method according to any one of claims 1 to 7.

Description

Method, device, equipment and medium for detecting target feature fusion enhancement of monitoring area Technical Field The invention relates to the technical field of monitoring area detection, in particular to a method, a device, equipment and a medium for detecting target feature fusion enhancement of a monitoring area. Background Monitoring area detection is the identification and monitoring of objects within a particular area using computer vision and image processing techniques. At present, the detection of the monitoring area mainly adopts a deep learning-based mode to realize target detection, and the target characteristics can be enhanced by fusing multi-scale characteristics so as to improve the target detection precision, but the mode has the following problems: 1. it is impossible to detect whether the target is newly added or not and the target area is changed In the conventional detection method for the target in the monitoring area based on the deep learning, the target is usually detected by using a deep learning model on a single frame image, but in an actual application scene, it is often required to determine whether the target in the monitoring area is newly increased or not and how large the area is newly increased. 2. It is difficult to accurately detect a small area target The traditional monitoring area detection method based on deep learning mainly comprises three implementation modes, namely a supervised method, a semi-supervised method and an unsupervised method, wherein the performance of the supervised method is the best, but the problem that the characteristic information of a small area cannot be effectively represented due to the fact that the scale of a deep network characteristic layer is small, the whole network model tends to be large in area and the area with obvious characteristic information, the problem that false detection and omission detection can occur in the small area is caused, and especially for small targets in the monitoring area in a complex scene, the small targets in the area are difficult to accurately detect. Disclosure of Invention Aiming at the technical problems in the prior art, the invention provides the monitoring area target feature fusion enhancement detection method, device, equipment and medium which are simple in implementation method, low in cost, high in detection precision and high in anti-interference performance, and can enhance the feature information of a small area and improve the detection precision and the anti-interference performance of the small target. In order to solve the technical problems, the technical scheme provided by the invention is as follows: A method for detecting target feature fusion enhancement of a monitoring area comprises the following steps: acquiring images of different time sequences of a monitoring area, inputting the images into a feature pyramid network to perform feature coding so as to extract features of different scales and perform feature fusion; Inputting a label image, and acquiring the center point coordinates of a target area, the area of which is smaller than a preset area threshold, in the label image; performing two-dimensional Gaussian fitting on the central point region according to the central point coordinates of the target region to generate small target region characteristic information; carrying out layered fusion on the characteristic information of the small target area and the characteristic information extracted by different scales respectively so as to fuse the characteristic information of the small target area to a corresponding decoding characteristic layer; and carrying out layered decoding and layered supervision on each decoding characteristic layer so as to judge whether the target in the target area of the monitoring area is newly increased or not and the change state of the target area according to pixel-by-pixel classification. Further, two-dimensional Gaussian fitting is performed on the center point coordinates of the target area by using a two-dimensional Gaussian function, and data in a specified range near the center point coordinates of the target area in data obtained by fitting is not 0, and data in other areas are 0. Further, two-dimensional Gaussian fitting is performed on the coordinates of the central point of the target area according to the following formula: Wherein, the Representing small target area characteristic information generated by fitting,,The pixel point position coordinates of the image are respectively,、For the preset coefficients of the two-dimensional gaussian function,,The coordinates of the center point of the target area are respectively obtained. Further, when a plurality of target areas with areas smaller than a preset area threshold exist, the small target area characteristic information generated by fitting each target area is combined to form final small target area characteristic information. Further, the step of carrying out layer