Extremely dark light object detection method based on RAW image

Abstract

The invention discloses an extremely dark light object detection method based on a RAW image, which is used for low-weak light images collected by a conventional image sensor. The method comprises the steps: establishing a dark light synthesis pipeline according to a physical imaging process, constructing a high-quality dark light object detection simulation data set based on existing normal light object detection data set resources, and training an accurate dark light object detection network. According to the method, extremely-dark light object detection based on existing common image acquisition equipment can be completed with high quality, high-efficiency and high-precision extremely-dark light object detection is realized while acquired images and human resources used for constructing an extremely-dark light object detection data set are remarkably saved, the detection precision is improved, the application scene of an object detector is expanded, and the bottleneck in the object detection field is broken. The invention can be used in multiple fields of deep space exploration, deep sea exploration, biomedicine, near-earth exploration and the like.

Description

technical field [0001] The invention relates to a method for generating dark light image detection data for extreme condition detection, in particular to a method capable of obtaining high-quality and high-fidelity extremely dark light images, and belongs to the technical field of computational vision. Background technique [0002] Extremely dark light object detection technology is a technology that can realize object detection tasks under low and weak light conditions, and can effectively detect target objects in dark light images with low brightness, obvious noise, and low signal-to-noise ratio. [0003] In low-low-light scenes with limited light sources, the image sensor receives fewer photons during the exposure time, which is limited by the physical characteristics of the image sensor, resulting in low brightness, obvious noise, and low signal-to-noise ratio in the captured image. The characteristics of the image seriously affect the information contained in the image....

AI Technical Summary

Problems solved by technology

One is to collect pairs of real data for learning and evaluation of the extremely dark light image detection network, which is similar to normal light image detection, but the collection and production of real low and low light data sets is costly, time-consuming, and difficult to label, making high-quality dark light The object detection data set requires a lot of manpower and material resources; the second is to generate real simulation data, which can not only make full use of the existing rich normal light object detection data resources, but also save a lot of data production costs, but the key lies in the simulation of synthetic low and low light images. Accuracy of Modeled Pipelines

Gaussian noise and Poisson noise are commonly used simulation noise models, however, in reality, dark light images generated under different photon levels contain more complex noise

In addition, the images captured by existing cameras and public image resources are almost all RGB format images processed by the built-in ISP algorithm, and the semantic information is lost. Due to the physical characteristics of the image sensor and the complexity of the imaging process, directly in the Simply adding noise or lowering the image brightness on the image of the existing normal light data set is not consistent with the noise generation process in the actual imaging process

Both inaccurate noise description and the way noise is added can have a huge negative impact on the quality of the simulation dataset

Images