CN-122024155-A - Freight monitoring method, system and product based on low-light image enhancement

CN122024155ACN 122024155 ACN122024155 ACN 122024155ACN-122024155-A

Abstract

The invention discloses a freight monitoring method, a freight monitoring system and a freight monitoring product based on low-light image enhancement. The method comprises the steps of firstly obtaining a freight channel monitoring image, converting the freight channel monitoring image into an HSV color space, judging whether the freight channel monitoring image is a low-light image through the brightness of a V channel, inputting the low-light image into a pre-trained light Zero-DCE++ enhancement model for real-time processing, training the model by adopting a no-reference loss function, generating pixel-level parameters, utilizing light enhancement curve iterative computation to output an enhanced image, and then inputting the enhanced or original image into a target detection model to complete identification, and uploading a final result to a cloud platform. The invention solves the problems of poor image quality, insufficient real-time edge deployment and weak suitability with downstream detection tasks of the existing enhancement algorithm in a low-light environment, and remarkably improves the accuracy, real-time performance and system reliability of night freight monitoring.

Inventors

MENG HU
ZHANG JIN
YANG YING
CHEN QING
LI YANG

Assignees

中国公路工程咨询集团有限公司
中交智运有限公司

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. A freight monitoring method based on low-light image enhancement, applied to Bian Yun cooperative systems, characterized in that the method is performed by an edge computing device, comprising: Acquiring an image to be processed in a monitoring video stream of a freight channel; converting the image to be processed into an HSV color space, and calculating the average brightness value of a V channel of the image; judging whether the average brightness value is lower than a preset brightness threshold value, if so, inputting the image to be processed into a pre-trained low-light image enhancement model to obtain an image after low light enhancement, otherwise, directly using the image to be processed, wherein the low-light image enhancement model is a pre-trained Zero-DCE++ image enhancement model; Inputting the obtained low-light enhanced image or the image to be processed into a pre-trained target detection model to identify and detect a freight target; and uploading the detection result to a cloud platform for visual management and data tracing.
2. The method according to claim 1, wherein if the brightness threshold is lower than the brightness threshold, inputting the image to be processed into a pre-trained low-light image enhancement model to obtain a low-light enhanced image, specifically comprising: Inputting the image to be processed into a pre-trained Zero-DCE++ image enhancement model to obtain an enhancement parameter diagram, wherein the Zero-DCE++ model is obtained through non-reference loss function training comprising space consistency loss, exposure control loss, color constancy loss and illumination smoothness loss; and carrying out iterative computation on the image to be processed through a light enhancement curve by utilizing the enhancement parameter map, and generating an enhanced image.
3. The method of claim 2, wherein the Zero-dce++ image enhancement model is model converted and quantized by RKNN tool-chain for deployment on an RK3588 chip-based embedded platform.
4. A method according to claim 2 or 3, wherein the running of the low light image enhancement model on the edge computing device is model-inferred by a neural network processor NPU and the post-processing computation of image enhancement is accelerated by a NEON instruction set.
5. The method of claim 2, wherein the light enhancement curve is calculated as: ; Wherein, the Representing the iteration number; Represent the first Post-iteration pixel Is a luminance value of (1); an enhancement parameter representing each pixel; representing the corresponding pixels output by the Zero-DCE++ model In the first place Enhancement parameters in the multiple iterations.
6. The method according to claim 2, wherein the Zero-dce++ model is trained by a no-reference loss function comprising spatial consistency loss, exposure control loss, color constancy loss, and illumination smoothness loss, specifically comprising: spatial consistency loss maintains the spatial consistency of an image by preserving the difference in neighboring regions between an input image and an enhanced image, as follows: ; Wherein, the Representing a spatial consistency loss; representing the number of local areas; expressed in terms of regions Y and I respectively represent average pixel values of local areas in the enhanced image and the input image; the exposure control loss is the distance between the average intensity value of the local area and the exposure good value E, and is expressed as follows: ; Wherein, the M represents the number of local areas, and D represents the average intensity value of the local areas in the enhanced image; a sequence number indicating a local area; representing the first in the enhanced image Average intensity values of the individual local regions; the color constancy loss is used to correct for possible color deviations in the enhanced image, creating a relationship between the three color RGB adjustment channels, as follows: ; Wherein, the Representing a loss of color constancy; Representing in an enhanced image The average intensity value of the channel is determined, Representing in an enhanced image Average intensity value of channels, a pair of channels is expressed as ; Representing the range of selecting paired channels, (R, G) representing a pair of channels R and G, (R, B) representing a pair of channels R and B, (G, B) representing a pair of channels G and B, R, G, B being the three color channels of red, green, blue constituting the image; The illumination smoothness loss is used to maintain a monotonic relationship between adjacent pixels, and one illumination smoothness loss is added to each curve parameter graph a, as follows: ; Wherein, the Indicating the loss of illumination smoothness, N indicating the number of iterations, N gradually adding to N with each 1-up in each round of accumulation, the horizontal and vertical gradient operations being respectively indicated as And ; R, G, B are the three color channels of red, green, blue that make up the image; A curve parameter diagram of a c channel in the nth iteration is represented; Total loss of Expressed as: ; Wherein the weight is And For balancing the effects of color constancy loss and illumination smoothness loss, respectively.
7. The method of claim 1, wherein the object detection model is a YOLOv s-OBB model for detecting and outputting an oriented bounding box of a shipping object.
8. A freight monitoring system based on low-light image enhancement is characterized by comprising a cloud platform and edge computing equipment, wherein, The edge computing device is deployed on the freight transportation channel site, and is used for executing the freight transportation monitoring method based on low-light image enhancement according to any one of claims 1-7 in real time, identifying abnormal transportation behaviors, and sending detection results and original data to the cloud platform; the cloud platform is used for receiving and storing data from the edge computing equipment and providing visual display, historical data query and tracing functions of detection results.
9. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the steps of the method according to any of claims 1-7.
10. A computer program product comprising a computer program or instructions which, when executed by a processor, carries out the steps of the method according to any one of claims 1 to 7.

Description

Freight monitoring method, system and product based on low-light image enhancement Technical Field The application relates to the technical field of image detection, in particular to a freight detection method, a freight detection system and a freight detection product. Background In logistics transportation systems in high-risk industries such as coal, chemical industry and the like, in-transit monitoring at night and in low-illumination environments such as tunnels and the like is a key link for guaranteeing transportation safety. However, the images collected by the traditional video monitoring system often have the problems of low brightness, poor contrast, obvious noise, loss of details and the like due to serious and insufficient ambient light, so that key visual information such as transportation vehicles, cargo states, surrounding environments and the like are difficult to effectively identify. The 'visual blind area' not only leads the performance of the traditional target detection algorithm depending on clear images to be reduced sharply, but also causes false alarm and missing alarm to occur frequently, and constitutes a great potential safety hazard. Therefore, developing a technology capable of actually improving the image quality in a low-light environment and supporting reliable intelligent analysis has become an urgent need in the field of freight safety management. Around low-light image enhancement, researchers at home and abroad have developed a great deal of work, and the main technical routes can be divided into a first category, namely, a physical model-based method, such as Retinex theory and variants thereof, which is sensitive to noise and complex in calculation by assuming that an image can be decomposed into illumination and reflection components and is optimized respectively, and is easy to generate halation artifacts and color distortion in a complex real scene, and a second category, namely, a deep learning-based method, such as supervised learning, is used for realizing end-to-end enhancement by utilizing paired data training depth networks, the performance of the deep learning-based method is seriously dependent on large-scale and high-quality training data, and the acquisition cost of low-light/normal-light image pairs is high, so that the generalization capability of the model is limited. Unsupervised or zero reference learning methods attempt to break away from the dependence on paired data, e.g. by designing special loss functions to constrain the enhancement results, but the resilience and stability to details under extremely low light conditions remain to be improved. Particularly, although the existing algorithm is continuously improved in laboratory evaluation indexes, the existing algorithm still faces a series of serious challenges in real industrial scenes, particularly in the ground of a freight monitoring system which needs front-end real-time processing, wherein most advanced enhancement models are high in computational complexity and large in parameter quantity, hard requirements of video stream real-time processing (such as >25 FPS) are difficult to meet on embedded edge equipment with limited computational power and power consumption, secondly, high-order visual tasks such as enhancement tasks and subsequent target detection are usually designed in an isolated mode and processed step by step, the enhancement process is guided by improving the subjective visual quality of images, and the feature which is critical to detection cannot be effectively reserved or enhanced in the enhancement process, so that the overall performance of the system cannot be optimized, and finally, the robustness and the adaptability of the existing algorithm are difficult to guarantee the stability of long-term running aiming at specific severe environments such as multiple illumination conditions and dust interference of freight channels. In summary, there is currently a lack of a low-light image enhancement solution that can compromise enhancement effect, computational efficiency, suitability for downstream tasks, and edge deployment feasibility, which is a core bottleneck that restricts the wide application of intelligent freight monitoring technology in low-light environments. Disclosure of Invention The application provides a freight monitoring method, a freight monitoring system and a freight monitoring product based on low-light image enhancement, and aims to solve the problems that in the prior art, the low-light enhancement algorithm has high calculation complexity, is difficult to meet the requirement of edge instantaneity, has insufficient suitability for downstream detection tasks and has poor deployment robustness in a complex industrial environment. In a first aspect, a freight monitoring method based on low-light image enhancement is provided, applied to Bian Yun cooperative systems, and the method is executed by an edge computing device and includes: Acquiring an ima