An Anti-aliasing Anchor-free Object Detector And Its Compression Method

With the increasing demand for object detection tasks in various fields,the challenges of object detectors are gradually diversifying.In order to solve the problem of aliasing and the lack of ideal results for multi-scale object,an Anchor-Free object detector is redesigned.In order to solve the problem of redundancy in the number of parameters and heavy computation in the feature learning network of Anchor-Free object detector,a new compression for Anchor-Free detector is proposed,further researches on Anchor-Free object detectors are carried out as follows:(1)An Anti-aliasing Anchor-Free object detectorTo address the problem of aliasing in the Anchor-Free object detectors,an anti-aliasing module is designed to filter out the useless high-frequency feature signals in the network channels by adaptively generated low-pass filters to avoid jaggedness in the down-sampling process and reduce the loss of image feature information in the spatial dimension.To address the problem that the Anchor-Free object detector is not fit for multi-scale objects,a multiscale feature fusion module is designed to extract the multi-scale feature information of the objects and fuse them by using atrous convolutions.The experimental results show that on Pascal VOC datasets,the m AP reaches 82.1%,and FPS reaches 32.The m AP and FPS are improved by 4.3% and 18.5% compared with the original Center Net-Resnet101 model.(2)An improved compression method for Anchor-Free object detectorsA double attention modules guided model compression method is designed to address the problems of large parameter redundancy,high computational overhead,and slow detection speed of the anchor-free object detectors.To compensate for the shortcomings of the original channel pruning algorithm,a lightweight and effective double attention module DAM(Double Attention Modules)is designed with the premise of reducing the complexity of the original attention model;under the guidance of the DAM,the channel pruning algorithm performs channel pruning on the anchor-free object detector with a new evaluation criterion,which reduces the number of parameters in the feature extraction network.The number of parameters of the feature extraction network is reduced,thus reducing the resource consumption of the model in terms of computation and storage,and improving the detection speed of the model.Experimental validation of the compressed Anchor-Free object detector was conducted on the commonly used datasets of PASCAL VOC,Image Net,and CIFAR-100.Taking the experiments on the PASCAL VOC dataset as an example,the results show that the number of Center Net-Res Net101 parameters before and after compression is reduced from56.95 M to 22.38 M and the FPS rises from 27 to 46 with only 0.6% loss of m AP.