Deep Model Compression In Computer Vision

Recent years,Convolutional Neural Networks(CNNs)have become the de-facto method in computer vision.Despite their remarkable effectiveness in a variety high-level vision tasks(such as classification and detection)and low-level vision tasks(such as neural style transfer,super-resolution),their practical applications are heavily contrained by the massive storage,computation,and energy consumption.M odel compression is an emerging area where researchers are devoted to removing the redundancy in deep neural network models for smaller storage,faster speed,etc.In this thesis,two structured pruning methods(SPP and IncReg)are proposed to accelerate modern convolutional neural networks in classificaition task.Meanwhile,in a low-level vision task,neural style transfer,the author also presents a model compression method(StyleDistill)to remove redundant filters in VGG19 based on knowledge distillation.Extensive experiments demonstrate the merits of the proposed methods compared with other state-of-the-arts.The specific novelties and contributions are summarized as follows:1.Parameter pruning is a kind of promising model compression methods,which aim for removing useless parameters of CNNs without serious perforamce degrade.Among them,structured pruning is to generate regular sparsity mainly for acceleration rather than storage reduction.Relative importance criterion is a critical however open tricky problem for pruning,which accounts a lot for the pruning effctiveness.Given the well-known difficulty to propose theretically sound and practically useful importance criteria,the author in this thesis claims to pay more attention to the pruning process design.In this spirit,SPP is proposed to assign a pruning probability to each weight in CNNs so that the pruning process can become softer,beneficial for network to adapt and recover during pruning.By adjusting these pruning probabilities,we can prune the weights gradually,and also we can correct the misjudgments caused by the lame importance criteria(such as L1-norm).SPP delivers encouraging results with column pruning on large CNNs(AlexNet,VGG16,ResNet50)on ImageNet dataset,but its training process is not so stable.To resolve this,the author further proposes IncReg,which is a regularization-based pruning method.IncReg adjusts the regularization factor before the L2 penalty term in an incremnetal manner.Same as SPP,it can also correct the relative importance misjudgments and be even softer,which is especially valuable when pruning a large number of parameters and pruning compact networks(e.g.,ResNet).IncReg reports even better results than SPP and many other state-of-the-arts.2.Apart from model compression in high-level vision,this thesis also focuses on the filed of low-level vision,where the complexity of large CNNs also hinders their applications.Specifically,in the task of neural style transfer,the author proposes StyleDistill method based on knowledge distillation to reduce the number of filters in VGG19.Based on an existing stylization algorithm WCT,the pruned model achieves comparable to even better stylization results with the original VGG19,albeit 1/15.5 times smaller.More impotantly,with the small model,we can achieve ultra-resolution(over 40 megapixels)universal style transfer for the first time on a single 12GB GPU.To show the generality of the proposed method,StyleDistill is also evaluated on photorealistic style transfer scenario(with PhotoWCT algorithm)and optimization-based artistic style transfer(with Gatys algorithm)with fairly pleasing results.