| Artificial intelligence(AI)will be becoming an important driving force for a new round of technological revolution,and as one of the most essential research contents in the direction of AI,object detection directly affects the subsequent image processing tasks,such as object tracking,pose estimation,automatic driving,etc.Object detection based on the deep neural network is the mainstream research direction in recent years,which can automatically extract the features of the multi-object in the image,then locate and identify the needed objects quickly and accurately.However,due to the angle,distance,complex scene,and light intensity conditions of the image shooting,it is inevitable that small objects and occluded targets appear in the image.In addition,they have a few effective pixels,carry a few features,are even not obvious,and are largely submerged in noise and dark background,which makes it hard to use the effective features.These reasons lead to the problems of false detection,missed detection,and repeated detection,thus affecting the detection accuracy.And with the exacerbation of network layers,it is easy to cause the loss of detailed features,so the detection of small objects and occluded targets has always been one of the most challenging problems in the field of object detection.To solve the above-mentioned questions,this paper conducts research on small objects and occluded targets detection algorithms based on deep regression neural network,and proposes a new method YOLO-ACN to detect small objects and occluded targets,and further based on the previous research work proposes YOLO-FIRI for the small objects and background occlusion problems in the infrared images.The main research innovations are summarized as follows:(1)To solve the problem that the object detection performance is poor in detecting small objects and occluded targets,a new object detection algorithm YOLO-ACN(Attention,CIo U,NMS)is proposed based on the deep regression neural networ.In the network design process,a two-dimensional attention mechanism is introduced to realize the attention of small objects and occluded targets,then use CIo U to calculate the boxes loss,and further use the CIo U as threshold in the post-processing stage to avoid the bounding boxes of small objects and occluded targets being filtered.(2)For the problem that low precision caused by long distances,weak energy,and low resolution in the infrared images,this paper further proposed an algorithm YOLO-FIRI for small objects and occluded targets in the infrared image.To effectively extract features,while designing the feature extraction network,the CSP(Cross Stage Path)module in the shallow layer is expanded and iterated.And an improved kernel attention module SK(Select Kernel Networks)is introduced in residual blocks to focus on objects.In addition,by improving the design of the multi-scale detection layer,the detection of shallow high-resolution feature maps is increased to improve the sensitivity of the small objects and occluded targets.(3)Given the problems of the poor resolution of infrared images and large background interference,an image fusion preprocessing method based on a deep neural network is used to fuse infrared images and visible images,then the quality of the infrared images is enhanced,the redundant information is reduced,the feature information is enriched.While realizing the data enhancement of infrared images,the feature extraction ability and detection performance of the model are further improved.Finally,this paper conducts some experiments on the natural image datasets MS COCO and VOC,and the infrared image datasets KAIST and FLIR.For the proposed model YOLO-ACN,the experimental results on the MS COCO show that the APs of small objects is improved by 1.0%,the speed is improved by7 ms,and the mean average precision(m AP50)reaches 53.8%.For the proposed YOLO-FIRI,the results on the KAIST dataset show that the m AP50 achieves 98.3%,with a real-time speed of 14 ms.On the detection of the class of people,the m AP50 even reaches 99.0%.The m AP50 on the FLIR was 83.5%.Compared with YOLOv5,the m AP50 of YOLO-FIRI on the two different infrared datasets is 5.2% and 3.4% higher,respectively.Moreover,the weight file size is only 15.0 MB,which is suitable for applications on embedded devices.To sum up,while maintaining the real-time speed of one-stage object detection,the proposed detection algorithms based on deep regression neural network can pay more attention to the detection of small objects and occluded targets. |
