Abstract:Low-light object detection is an important research area in computer vision, but it is also a difficult issue. This research offers a low-light target detection network, NLE-YOLO, based on YOLOV5, to address the issues of insufficient illumination and noise interference experienced by target detection tasks in low-light environments. The network initially preprocesses the input image with an improvement technique before suppressing high-frequency noise and enhancing essential information with C2fLEFEM, a unique feature extraction module. We also created a multi-scale feature extraction module, AMC2fLEFEM, and an attention mechanism receptive field module, AMRFB, which are utilized to extract features of multiple scales and enhance the receptive field. The C2fLEFEM module, in particular, merges the LEF and FEM modules on top of the C2f module. The LEF module employs a low-frequency filter to remove high-frequency noise; the FEM module employs dual inputs to fuse low-frequency enhanced and original features; and the C2f module employs a gradient retention method to minimize information loss. The AMC2fLEFEM module combines the SimAM attention mechanism and uses the pixel relationship to obtain features of different receptive fields, adapt to brightness changes, capture the difference between the target and the background, improve the network's feature extraction capability, and effectively reduce the impact of noise. The AMRFB module employs atrous convolution to enlarge the receptive field, maintain global information, and adjust to targets of various scales. Finally, for low-light settings, we replaced the original YOLOv5 detection head with a decoupled head. The Exdark dataset experiments show that our method outperforms previous methods in terms of detection accuracy and performance.

Low-Illumination Image Enhancement for Foreign Object Detection in Confined Spaces

Research on Low-Light Image Enhancement Algorithm Based on Filtering and Data-Driven Composite

Reactive oxygen species production and antioxidant enzyme activity during epididymal sperm maturation in Corynorhinus mexicanus bats.

A dark image enhancement method based on multiscale features and dilated residual networks

EEMEFN: Low-Light Image Enhancement Via Edge-Enhanced Multi-Exposure Fusion Network

LLOD:A Object Detection Method under Low-Light Condition by Feature Enhancement and Fusion

A Low-Light Image Enhancement Algorithm Incorporating Cross-Mixed Attention and Receptive Field Expansion Mechanism

IRNet: an Improved Zero-Shot Retinex Network for Low-Light Image Enhancement

You Only Need One Color Space: An Efficient Network for Low-light Image Enhancement

Event-Based Low-Illumination Image Enhancement

A novel low light object detection method based on the YOLOv5 fusion feature enhancement

Zero-Shot Enhancement of Low-Light Image Based on Retinex Decomposition

Exploring Image Enhancement for Salient Object Detection in Low Light Images

Low-light-level image enhancement algorithm based on integrated networks

DarkFeat: Noise-Robust Feature Detector and Descriptor for Extremely Low-Light RAW Images.

Dynamic Low-Light Image Enhancement for Object Detection Via End-to-End Training.

Deep Bilateral Retinex for Low-Light Image Enhancement

A Two-Stage Unsupervised Approach for Low Light Image Enhancement

Learning to Enhance Low-Light Image via Zero-Reference Deep Curve Estimation

EMNet: Edge-guided multi-level network for salient object detection in low-light images

Enhancing Low-Light Images Using Infrared-Encoded Images