Research On Human Object Detection Method Based On Image Compression Denoising

Object detection is one of the research directions of computer vision,which has important research and application value in the fields of unmanned driving,intelligent transportation,security monitoring,human-computer interaction and so on.The current development of deep learning technology has greatly promoted the progress of object detection algorithms.However,human object detection due to object scale changes,where are still many problems with detection methods due to changes in attitude,occlusion,and complex background.In addition,in order to efficiently transmit a large number of images and reduce network load,various social and video software will perform lossy compression on the original image,resulting in the loss of many image features,adversely affect subsequent object detection.In response to the above problems,this paper uses deep learning technology to conduct in-depth research on image compression noise removal and human object detection.A human object detection method based on image compression noise removal is proposed,and high-performance results are obtained on PASCAL VOC human body data sets.The main research contents are as follows:1.A method for noise removal in human body image compression based on deep convolutional network is proposed: unlike previous method that only uses RGB three channels,this paper focuses on available information of quality factor when JPEG algorithm compresses the image.By improving ResNet18 fully connected layer and loss function,a quality factor estimation method based on deep learning Res Net network is proposed to provide more reference information for removal of compressed noise.FFDNet network is used to combine tensor formed by the pixel?shuffle inverse transformation of image and tensor formed by quality factor,which is spliced by channel,and the image is compressed and noise removed.Experimental results show that the proposed method enables network to use additional information to effectively improve the quality of human body image compression noise removal,and provide support for the subsequent improvement of human object detection performance.2.A human object detection method based on high-resolution feature fusion is proposed: in view of insufficient features extracted by deep network,lacking of robustness under the influence of variable object environment and scale changes,and misjudgment of a single object detector,and take the human body image with compressed noise removed as input,and HRNet is used as the backbone network to extract high-resolution features,and combined with HRFPN to fuse low-level apparent information and high-level semantic information to obtain high-quality feature representations of object.Using cascaded object detectors for classification and bounding box regression to make inferred object box more accurate and further improve object detection performance.Experimental results show that the image after noise removal has better detection performance than the original human body image without noise removal,regardless of intuitive perception or specific quantitative indicators.Especially when dealing with multi-scale and different posture human objects,the detection effect is significant.