Combined Image Enhancement for Recyclable Waste Object Detection In Low-Light Environment

An improved DETR target detection algorithm is proposed to address the problems of low accuracy and high false detection rate of detecting recyclable garbage by traditional target detection algorithms under low light environment. Firstly, the low illumination dataset is synthesized using ganuna transform and Gaussian noise on the homemade recyclable garbage dataset to simulate the low illumination environment. Secondly, the ResNet50 network of the DETR model is replaced by the ResNeXt50 network, which improves the accuracy by a small margin, and the mAP0.5 is improved by 1.6%. Finally, to solve the problem of poor detection effect of the algorithm in low-light environment, the noise removal module and low-light enhancement module are added before the feature extraction network to significantly improve the detection performance of the DETR algorithm in low-light environment. The experimental results show that the average accuracy is improved by 25.5% to 96.4% compared with the conventional DETR model. The performance is also excellent compared with mainstream target detection models. When detecting recyclable garbage in a low-light environment, the algorithm can perform the detection task better and shows comparable detection results to those in normal environments.