Research On Target Detection Algorithm With Deformable Convolution Characteristics

With the rapid development of science and technology,the performance of computer hardware has been continuously improved,which has promoted the progress and development of computer vision.Research directions such as target detection,target segmentation,and target tracking have received extensive attention from researchers and many technology companies,and target detection is the basis of some of these research directions,so the importance of object detection research is self-evident.However,in practical application scenarios,there are some difficulties that restrict the performance of target detection algorithms.First of all,the differences in target types,the rich actions of flexible targets,and the angle of image acquisition will all cause large changes in the shape and scale of the detected target.Secondly,the distance from the image acquisition device or the small size of the target itself will lead to less pixels of the target.To solve the above problems,this paper proposes a target detection algorithm with deformable convolution characteristics.The specific work of this paper is as follows:(1)For the problem of changes in the shape and scale of target.First,since the shape and size of the conventional convolution kernel are fixed,it cannot cope with the shape and scale changes of the target.In order to solve this problem,the deformable convolution kernel is introduced.Compared with the conventional convolution kernel,the deformable convolution kernel adds horizontal and vertical offsets to each point.The shape of the convolution kernel can be adaptively adjusted according to the shape and scale of the target,so that the receptive field of the convolution kernel can focus on the target object to achieve better feature extraction effect.Secondly,the calculation of the specific implementation process of deformable convolution is optimized.Since the offset of the deformable convolution kernel is the number of floating points,bilinear interpolation is needed to calculate the pixel value of the offset position.The convolution operation of multiple convolution kernels in the feature map at the same level will produce a large number of repeated calculations.Therefore,this paper conducts preprocessing before the deformable convolution calculation,calculates the pixel value of the offset position according to the offset,and caches it to generate a new feature map to avoid the subsequent redundant calculation.Finally,in order to achieve the optimal balance between accuracy and speed of the improved algorithm,three progressive deformable convolution kernel embedding schemes are designed for comparison,and the optimal scheme is selected for subsequent improvement.(2)For the problem of small target detection.First,the feature pyramid network is analyzed.The feature pyramid network transmits the rich semantic information of the deep feature map downward through the top-down path and fuses the detailed feature information of the shallow feature map for the final prediction.In order to make full use of the shallow feature information,this paper proposes a channel feature enhancement method with feature protection mechanism.Under the premise of ensuring that the feature information is not lost,the shallow features are enhanced to participate in the final feature fusion process.Secondly,because the deep features involved in feature fusion are directly up-sampled and then fused with shallow features,this process does not include learnable parameters,which will cause spatial dislocation problems when fusing features at different levels.Therefore,this paper performs feature alignment operations on deep features before feature fusion,uses convolution calculations to learn the spatial position errors of deep and shallow features,and completes the adjustment of deep features through deformable convolution calculations.