Research On Low Illuminance Object Detection Methods Based On Transfer Learning

In today’s social production and life,computer vision technology has been applied to practice in many fields,such as video surveillance,robot detection and uav reconnaissance.At present,a lot of researches have been done on normal illumination object detection technology,but there are few researches on low illumination image object detection.Images taken in scenes with insufficient illumination or uneven illumination generally have problems such as low illumination,insufficient contrast and serious loss of detail information.These problems make low-illumination images unable to contain enough information,at the same time,low-illumination image capturing and object labeling are difficult,which requires a lot of time and effort.These problems are not conducive to the construction of large-scale low-illumination object detection datasets.In this paper,the subject of low-illuminance image object detection is studied.By studying the image feature enhancement and crossdomain object detection technology based on transfer learning,the problems of insufficient information and scarce data samples of low-illuminance images can be solved,so that lowilluminance images can also be the same as normal illuminance images,with good detection efficiency.Using transfer learning as a key technology,this paper proposes the following two low-illuminance object detection methods for the above different scenarios:Aiming at the problem that low illumination object detection accuracy is seriously affected by uneven illumination and low contrast in low illumination image.A low illumination object detection method based on feature enhancement and multi-scale receptive field(FEMR)is proposed.First,the pixel-level high-order mapping module is used to learn the high-order mapping relationship between low illuminance and normal illuminance,so as to improve the significance of low-illuminance object features and obtain preliminary enhanced feature information.Then,the key information enhancement module combined with various attention mechanisms was used to highlight important features and filter noise information to obtain further enhanced feature information.In addition,the long distance feature capture module is used to introduce strip receptive fields of various scales to capture the long distance relationship of isolated regions in low illumination scenes.Experimental results show that FEMR has a good performance in low-illuminance object detection accuracy,and can directly output the detection results under normal illuminance style image to achieve end-to-end low-illuminance object detection,which is convenient for human eyes to directly evaluate the accuracy of detection results.In view of the scarcity of low-illuminance image samples when deep learning method is used to train low-illuminance object detector,a method based on Faster R-CNN object detector.A low illuminance object detection network framework based on Faster R-CNN object detector combined progressive alignment and prototype alignment(CPA-NET)is designed to use large-scale normal illuminance image datasets are used to supplement detection information for small-scale low-illumination image data to achieve unsupervised low-illumination object detection task.Firstly,the image-level and instance-level domain progressive alignment modules are used to make the detector unable to distinguish the domain of the image.Consistency regularization loss is used to ensure the consistency of domain classification results at both levels.At the same time,the information aggregation module based on graph convolution is used to complete the prototype representation of each category.On this basis,a re-weighting module that balances the proportion of samples is used to assign a higher weight to the scarce samples to coordinate the domain adaptation process among different classes.The experimental results show that the CPA-NET lowilluminance object detection method constructed in this paper has better detection performance than the current advanced cross-domain object detector,and can make the object detector have the object detection ability of the whole scene.