Research On Road Garbage Detection Method Based On Adaptive Low-light Images Enhancement

It is challenging to efficiently perform high-precision inspection tasks in complex natural environments.Machine vision technology is widely used in many fields,such as autonomous driving and smart city construction as an effective method to accomplish this task.At present,the detection technology that is applicable to well-lit scenes is more mature.However,as for the task of detecting small targets in low-light scenes,due to the captured images with low dynamic range,as well as some interference factors such as color bias and noise,the relevant detection algorithms cannot accurately locate the targets.In this paper,a fast enhancement algorithm for low-light natural scenes,as well as its paired and unpaired training methods are proposed,in order to handle the problems of poor algorithm adaptability,high training cost,and low detection accuracy and efficiency in low-light scenes.The proposed enhancement algorithm is combined with the detection algorithm to realize the fast low-light garbage recognition in a low-cost experimental platform.The main contributions of this paper are as follows:(1)A two-stage low-light image adaptive enhancement method based on efficient feature aggregation is proposed.Aiming at the problem of poor adaptability of low-light image enhancement algorithms to different scenes,a two-stage Feature Aggregation Enhancement Net(FAEN)is designed to realize the brightness adjustment and detail correction sequentially.A global feature extraction module is introduced to enhance the multi-scale feature extraction capability of the dual Unets backbone network.And a learnable regularized attention module is proposed to balance the enhancement effect of different regions.An adaptive feature aggregation module is also used to shorten the feature path between two stages,making the enhanced image contain richer detail information.The results of comparison experiments on several paired datasets show that the proposed FAEN is more adaptable to different scenes,and can effectively suppress the noise,color bias and artifacts generated during the enhancement process.The PSNR index reaches 31.826 in the self-built multi-scene paired low-light road dataset,which represents a significant improvement compared with the stateof-the-arts.In addition,after being deployed in the embedded Jetson AGX Xavier edge computing platform with 30 W power consumption,the optimized algorithm achieves the enhancement speed of 22.78 frames per second for low-light images with 2432×896 pixels.(2)A GAN-based adversarial training method for low-light image enhancement network is proposed.To solve the problems of the difficulty of capturing paired images in practical working scenarios and the high-cost paired training,the adversarial training with unpaired images is proposed.A domain adaptive adversarial generative network is adopted to calculate the global adversarial loss,which helps the enhancement algorithm study the mapping from the low-light feature domain to the normal-exposure feature domain.Besides,a gammacorrection-based self-supervised content loss is designed to avoid texture and structural distortion in the generated outputs.In order to enhance the local detail retention and noise suppression capability,a Patch GAN-based local unsupervised noise reduction module is additionally introduced.Experimental results show that the adversarially trained enhancement model exhibits stronger generalization ability in multiple publicly unpaired low-light datasets with an average BRISQUE index of only 17.58,which is better than the comparison algorithms.(3)A model-fusion-based low-light garbage detection method for low-light scenes is proposed,which can efficiently identify garbage targets in the edge computing platform.Due to it’s difficult for the detection algorithm to extract features from low-light images effectively,an additional feature extraction module of the enhancement algorithm is integrated into the detection algorithm,which can fuse features with a flexible structure,thus improving the detection performance.In addition,in the embedded simulated lighting experimental platform built on the basis of the actual working scene,the model-fusion-based low-light garbage detection method is evaluated.The results show that it can achieve a detection rate of 0.851 and an inference speed of 10.80 frames per second for low-light images of 2432×896 pixels in the embedded platform.