Object Detection Research On Multi-band Images

Target detection technology is one of the core technologies in security,road traffic monitoring and other systems.By extracting key target features in the image,the category and location information of the target are analyzed.Traditional target detection algorithms based on visible light band images have made breakthroughs in small-scale target detection and dense target detection,and their effectiveness has been confirmed by large-scale natural image datasets.However,under the conditions of insufficient illumination and smoke occlusion,due to the loss of effective target information,it is still difficult for such algorithms to obtain relatively robust detection results,which seriously affects the overall reliability of the all-weather monitoring system.In this paper,the visible light sensor and the far-infrared sensor are combined,and a target detection algorithm based on multi-band images is designed,and the multi-band image information with high registration quality collected by the two sensors which is reasonably used to improve accuracy in complex scenes.Due to the fact that there are few researches on target detection algorithms based on multi-band images,and most of them only analyze pedestrian targets due to the limitation of data.This paper firstly optimizes the pedestrian detection algorithm in infrared images to verify the effectiveness of using multi-band images information.Secondly,it analyzes the overall fusion framework and fusion strategy in the process of multi-band image feature-level fusion,and optimizes the learning of network feature fusion.Finally,through the selfbuilt multi-band data set and experimental analysis,it effectively proves that the algorithm in this paper is advanced in detection accuracy and inference speed in multiclass target detection.The specific work of this paper is shown as follows:1)Research on pedestrian detection algorithm based on infrared images.Because infrared images can better overcome the influence of scene illumination and there are few related research theories,this paper first uses infrared band images as the information source of target detection,and designs a multi-task learning framework to enhance network performance.Experiments have effectively confirmed that the detection performance of the algorithm is better than the current infrared pedestrian detection algorithm,but there is still a large accuracy gap compared with the mainstream algorithms using multi-band image information,which confirms the effective value of this topic.2)Research on pedestrian detection algorithm based on multi-band images.In view of the structural redundancy of mainstream multi-band pedestrian detection algorithms,the single fusion scale and the limitations of traditional fusion strategies,this paper conducts research on the overall fusion framework and feature fusion strategy.A three-branch mid-term fusion model is constructed as the overall framework of this paper,and a gated network is designed to sense the global information of the image and gradually guide the fusion of local information.The experimental results effectively confirm that the algorithm proposed in this paper can achieve better detection performance than similar mainstream algorithms at a high inference speed.3)Research on multi-class target detection algorithm based on multi-band image.Aiming at the lack of multi-class target information of multi-band images,this paper designs a target labeling tool for multi-band images,and builds a multi-class multibands object detection datasets through data collection and sorting with adaptive histogram equalization processing and other improvement.In addition,for the problem of long-tail distribution of target samples,the balanced Focal loss function is used to strengthen the attention of network to the few-sample target,and experiments on the self-built data set confirm that the optimized algorithm achieves impressive detection result with high accuracy and fast speed.