CN-122023763-A - Weak light domain self-adaptive enhanced target detection method based on zero sample

CN122023763ACN 122023763 ACN122023763 ACN 122023763ACN-122023763-A

Abstract

The invention discloses a zero-sample-based weak light domain self-adaptive enhanced target detection method which comprises the steps of constructing a network architecture comprising a decoder based on a target detector, introducing a pretrained Retinex decomposition network as an aid, pairing a synthesized dim light image and a good illumination image to serve as network input, splitting a backbone network according to the Retinex theory, decoding reflectivity by utilizing shallow features, designing mutual feature alignment loss to strengthen illumination invariance learning, decomposing the image into reflectivity and illumination based on the Retinex theory, exchanging the reflectivity to reconstruct the image, then decomposing again, and freezing pretraining weights of the Retinex decomposition network. The method has the beneficial effects that the technical scheme realizes the end-to-end joint optimization of image enhancement and target detection by fusing the multi-task targets such as detection loss, characteristic alignment loss, decomposition loss and the like through the total loss function, avoids the limitation of independent design of a preprocessing and detection module in the traditional method, and improves the overall performance of the system.

Inventors

DING HAIQIN
LI NING
CUI LINGLING
Qin Tongchun
JIA GUODONG
JI GUANG
LI ZHIWEI

Assignees

南通理工学院
江苏优众微纳半导体科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251223

Claims (10)

1. A weak light domain self-adaptive enhanced target detection method based on zero samples is characterized by comprising the following steps: S1, constructing a network architecture containing a decoder based on a target detector, introducing a pretrained Retinex decomposition network as an aid, and pairing a synthesized dim light image and a well-lighted image to serve as network input; S2, splitting a backbone network according to a Retinex theory, decoding reflectivity by utilizing shallow features, and designing mutual feature alignment loss to strengthen illumination invariance learning; S3, decomposing the image into reflectivity and illumination based on a Retinex theory, exchanging the reflectivity to reconstruct the image, decomposing again, and introducing re-decomposition consistency loss to enhance the reflectivity to represent learning; And S4, freezing the pre-training weight of the Retinex decomposition network, training and optimizing the network architecture, and adjusting the weight through a total loss function to improve the dim light detection performance.
2. The method for detecting an enhanced target of zero-sample-based low-light-field adaptation as claimed in claim 1, wherein in the step S1, the network architecture is DAI-Net, the decoder is a reflectivity decoder, and the Retinex decomposition network is frozen during training, and is only used for deducing the reflectivity and illumination false-true values of the input image.
3. The method for detecting a target with enhanced light field adaptation based on zero samples as claimed in claim 1, wherein the step S1 is characterized in that the composite dim light image is generated by a low-light synthesis pipeline based on physical heuristics, and is paired with an original well-lighted image for input.
4. The method for enhanced target detection based on zero-sample weak light domain adaptation of claim 1, wherein the splitting of the backbone network in step S2 is performed by splitting at a second convolution layer of the detector backbone network, the first half being denoted as For extracting low-level features and connecting to a reflectivity decoder, the whole backbone network being denoted as For the target detection task.
5. The method for detecting a weak light domain adaptive enhancement target based on zero samples as set forth in claim 4, wherein said reflectivity decoder is a lightweight module composed of two Conv+ReLU layers and is shared with the detection head The extracted illumination invariant features.
6. The method for enhanced target detection based on zero-sample weak light domain adaptation of claim 1, wherein the mutual feature alignment loss in step S2 The expression of (2) is: Wherein, the Represents the degree of divergence of KL, And The light well-lighted image and the synthetic dim light image are respectively passed through The characteristics of the output.
7. The method for detecting a target with enhanced light field adaptation based on zero samples as claimed in claim 1, wherein the image decomposition in step S3 is specifically performed by combining a dark-light image Is decomposed into reflectivity And illumination Good-illumination image Is decomposed into reflectivity And illumination 。
8. The method for adaptively enhancing a target detection in a weak light domain based on zero samples as set forth in claim 7, wherein the method for reconstructing an image by using the interchangeable reflectivity in step S3 is as follows: by means of And (3) with Reconstructing a dim light image ; By means of And (3) with Reconstructing well-illuminated images 。
9. The method for enhanced target detection based on zero-sample weak light domain adaptation of claim 8, wherein said re-decomposition of consistency loss in step S3 The expression of (2) is: Wherein, the And Respectively reconstructed images And Reflectivity decomposed by the reflectivity decoder.
10. The method for enhanced target detection based on zero-sample weak light domain adaptation of claim 1, wherein the total loss function in step S4 The expression of (2) is: Wherein, the In order to detect the loss of the material, Representing mutual feature alignment loss; representing a re-decomposition consistency loss; representing a reflectivity learning loss; Representing image decomposition loss; A weight representing mutual feature alignment loss; Weights representing re-decomposition consistency loss; Wherein, the reflectivity is learned and lost And image decomposition loss Can be expressed as: Wherein, the Representing the reflectivity decoder output; Representing a false true value; representing the average absolute error between the two, Representing a structural similarity index; representing image reconstruction loss; Indicating an illumination smoothness loss; Representing a constant reflectance loss; And Representing the weight of the corresponding loss.

Description

Weak light domain self-adaptive enhanced target detection method based on zero sample Technical Field The invention relates to the technical field of target detection, in particular to a weak light domain self-adaptive enhanced target detection method based on a zero sample. Background Target detection is a core topic in the field of computer vision, aimed at accurately identifying and locating target objects in images. In recent years, with the advent of large-scale labeling data sets such as COCO and Open Images and the development of advanced algorithms such as YOLO and fast R-CNN, the accuracy and efficiency of target detection under a sufficient illumination scene have been remarkably improved. However, in low-light or dim-light environments (such as nighttime, low-light indoor), the performance of the existing detector is greatly reduced due to the problems of low visibility, color distortion, noise interference and the like of the image, and practical application requirements (such as nighttime security monitoring, automatic driving nighttime environment sensing and the like) are difficult to meet. In order to solve the target detection challenge in the low light scene, researchers put forward two main schemes, namely image enhancement preprocessing, namely enhancing the low light image quality based on Retinex theory, deep learning and other methods, and then inputting the low light image quality into a detector for recognition. The Retinex theory adjusts the illumination component to enhance the image visibility by decomposing the image into reflectivity (object-inherent properties, illumination invariant) and illumination component (ambient illumination, variable). However, such methods rely on a large number of pairs of low-light-normal-light image data for training, and the low-light data is high in acquisition and labeling cost and limited in scene coverage, resulting in limited generalization capability. Detector domain adaptive tuning-a small amount of low-light data tuning model is used on the basis of a pre-trained detector on normal-light data. However, the problem of "domain offset" (difference in feature distribution) of low-light scenes is prominent, fine tuning tends to result in overfitting, and still requires low-light annotation data of a certain scale, failing to achieve "zero-sample" adaptation. In addition, the prior method independently processes image enhancement and target detection, and does not fully utilize the inherent correlation between the image enhancement and the target detection, wherein the enhanced image possibly introduces artifacts and interferes with detection, and a detector does not explicitly learn the characteristic representation robust to illumination change. Therefore, how to improve the self-adaptive capacity of the model to the low-light environment by fusion image enhancement and joint optimization of detection tasks under the condition of no low-light annotation data becomes a technical problem to be solved currently. Aiming at the problems, the invention provides a zero-sample-based weak light domain self-adaptive enhanced target detection method, which can realize improvement of detection performance in a weak light scene without depending on real low-light annotation data by constructing a collaborative learning framework. Disclosure of Invention The invention mainly solves the technical problem of providing a zero-sample-based weak light domain self-adaptive enhanced target detection method, which solves one or more of the problems in the prior art. In order to solve the technical problems, the invention adopts a technical scheme that the method for detecting the enhancement target based on the weak light domain self-adaption of the zero sample is characterized by comprising the following steps: S1, constructing a network architecture containing a decoder based on a target detector, introducing a pretrained Retinex decomposition network as an aid, and pairing a synthesized dim light image and a well-lighted image to serve as network input; S2, splitting a backbone network according to a Retinex theory, decoding reflectivity by utilizing shallow features, and designing mutual feature alignment loss to strengthen illumination invariance learning; S3, decomposing the image into reflectivity and illumination based on a Retinex theory, exchanging the reflectivity to reconstruct the image, decomposing again, and introducing re-decomposition consistency loss to enhance the reflectivity to represent learning; And S4, freezing the pre-training weight of the Retinex decomposition network, training and optimizing the network architecture, and adjusting the weight through a total loss function to improve the dim light detection performance. In some embodiments, the network architecture in step S1 is DAI-Net, the decoder is a reflectivity decoder, and the Retinex decomposition network is frozen during training, only to infer reflectivity and illumination artif