| 摘要: |
| 目标检测技术旨在从图像或视频中对特定类别目标进行定位和识别。然而，在低照度场景中面临对比度低、边界模糊、噪声干扰等问题，导致检测算法性能下降。为此，提出了一种基于颜色通道变换增强的目标检测（Color Channel Transformation Enhancement-based Object Detection，C2TEOD算法。首先，构建了颜色通道变换模块，引入可学习参数对不同颜色通道进行变换，提升增强策略灵活性。然后，通过一个图像增强模块对图像进行预处理，并采用检测损失联合优化增强模块与检测网络，使增强网络能够朝着有利于检测任务的方向优化。此外，提出了选择性自监督回归损失，利用原始低照度图像和增强图像输入检测网络并对其进行优化，再根据它们的检测结果对增强模块进行自监督回归优化，进一步提升检测效果。实验结果显示，该算法相比基线方法在Exdark、M3FD、LLVIP数据集上的平均精度均值（mean Average Precision，mAP分别提升2.2%、1.1%和0.2%。 |
| 关键词: 目标检测 图像增强 深度学习 联合优化 自监督 |
| DOI：10.20079/j.issn.1001-893x.250506001 |
|
| 基金项目: |
|
| Color Channel Transformation Enhancement-based Low-illumination Images Object Detection |
| ZHANG Min,QIAO Wensheng,ZHU Peipei,ZHU Sihan,ZHAN Yufei,HUANG Xiaochen,CHEN Honggang |
| (1.National Key Laboratory of Complex Aviation System Simulation，Southwest China Institute of Electronic Technology,Chengdu 610036;2.College of Electronics and Information Engineering， Sichuan University，Chengdu 610065，China) |
| Abstract: |
| Object detection technology aims to locate and identify specific category targets in images or videos.However,in low-illumination scenarios,problems such as low contrast,blurred boundaries,and noise interference,result in the decline of detection performance.To address this,a Color Channel Transformation Enhancement-based Object Detection(C2TEOD algorithm is proposed.Firstly,a color channel transformation module is constructed,and learnable parameters are introduced to transform different color channels,enhancing the flexibility of the enhancement strategy.Then,an image enhancement module is employed to preprocess the input images.This module is jointly optimized with the object detection network using detection loss functions,thereby enabling the enhancement module to learn to generate representations that explicitly facilitate the subsequent detection task.Additionally,a selective self-supervised regression loss is proposed that uses both the original low-illumination images and the enhanced images as inputs to optimize the detection network.According to detection results,the enhancement module is further optimized through self-supervised regression to improve detection performance.Experimental results show that,compared with the baseline method,the mean average precision（mAPmetrics on the Exdark,M3FD,and LLVIP datasets are improved by 2.2%,1.1%,and 0.2% respectively. |
| Key words: object detection image enhancement deep learning joint optimization self-supervision |
