Research And Application Of Target Detection Algorithm Based On Convolutional Neural Network

Object detection is one of the main tasks of computer vision and has a wide range of applications.With the rapid spread and layout of applications such as intelligent video surveillance,driverless technology and smart medical care,the detection of significant targets in image sequences has been extensively and intensively studied.However,the detection of dynamic non-significant targets such as occlusion is more serious,dense and small is still a challenging technology,which has become the bottleneck of computer vision to practical application.There are still some key technologies to be broken,including: Fraud,real-time and stability;how to quickly and accurately detect a specific area and its boundaries,for applications such as unmanned vehicles,intelligent robots,and drones that operate in large areas or irregular areas.It is also a major problem in object detection technology.For the anti-fraud problem,for the human face detection application,with the difference of the 3D and 2D face distribution of near-infrared light,an iterative quadratic frame difference model is proposed to quickly extract the living face and the non-living person with significant differences.Face features,then create a universal subspace template for PCA for live face detection.The average time taken to detect one frame of image under the condition of 2.3GHz CPU hardware is 1.57ms;it is tested in a near-infrared self-built database of 18 people.When the face is facing the near-infrared light source and looking up,the detection rate is 100%;When the right side or the left side is 30°,the detection rate is 98%;when the head is 30°,the detection rate is 97%.The experimental results show that the proposed human face detection model has rapidity and high detection accuracy.For the real-time and stability problems of target detection,in order to improve the accuracy and speed of face detection in a single image,improve the Mobile Net basic network with less parameters and high precision,and add a combination of eight standard convolutional layers,BN and Re LU6.Used to improve the characterization capabilities of the network.Based on this,the bottom-to-top structure of the feature pyramid structure is constructed,and the Mobile Net-SSD(MS)fast face detection model is established.Aiming at the rapid and stable detection of faces in image sequences,the MS model is merged with the Hog-based kernel correlation filtering algorithm to construct the MS-KCF model.Under the GPU GTX1080 hardware condition,the recall rate of the Easy,Medium,and Hard test subsets of the MS model in the WIDER FACE public data benchmark was 93.11%,92.18%,and 82.97%,respectively.The recall rate of the face data reference in the FDDB was published.93.60%,the average test speed is 84fps;the MS-KCF model is tested in the OTB data benchmark with excessive angle change and occlusion of the Girl and Face Occ1 image sequences.The recall rates are 84.3% and 94.6%,respectively,and the test speed is 193 fps.The experimental results show that the MS model and MS-KCF model proposed in this paper are fast,stable and have high detection accuracy.For the large-area or irregular area identification and its boundary detection problem,a fast and accurate PULNet semantic segmentation network and an eight-neighbor code traversal method to locate the boundary position of the pixel values ??in the region are proposed for the application of turf boundary detection.Validated in the ADE20 K public data benchmark,with an average regional coincidence of 32.86% and an average pixel accuracy of 75.65%,for 512 ×512 images,the average test speed under GPU GTX 1080 Ti hardware is82.7 fps;Self-built lawn database and field test under different environmental and weather conditions,with an average area overlap of 96.32%,for 848×480 size pictures,the average speed tested under GPU GTX 1080 Ti hardware is 67.3fps,on GPU GTX1050,CPU The average test speed of the I5-7300 hq laptop is 30.3fps.The experimental results show that the PULNet semantic segmentation network and the eight-neighbor code traversal method proposed in this paper have higher boundary detection accuracy and faster speed.