Research On Lightweight Salient Object Detection Method

Inspired by the human attention mechanism,salient object detection(SOD)aims at separating the salient object region from the background region in an image,which can be used as a preprocessing step for downstream tasks in the field of computer vision.Most saliency detection models pay too much attention to model accuracy,resulting in a large number of model parameters and slow inference speed.Most saliency detection models usually have tens of millions or even billions of parameters.Such complex and cannot be applied to scenarios with real-time requirements or limited computing resources.In order to solve the above problems,this thesis conducts research on lightweight models designed on two devices with different computing power,and uses pruning algorithm and quantization algorithm to further compress the model.The main work is as follows:(1)For devices with strong computing power,a novel convolutional neural network model based on top-down decoding flow-based framework is proposed,namely densely nested topdown decoding flow.In this framework,each decoding stage aggregates all higher level encoding features with the help of progressive compression shortcut path to make full use of the high level features and to alleviate the gradient vanishing problem.The densely nested top-down decoding flow can cooperate with lightweight backbone using a high information compression ratio,so that a backbone network equipped with a small number of additional parameters can contribute to high salient object detection performance.(2)For low computing power devices,this thesis proposes a simple and effective lightweight salient object detection network(Simple-SOD)from scratch.Unlike previous work that focused on neural network architecture search and employed complex convolutional modules,Simple-SOD adopts self-designed backbone,a simple feature fusion method and a efficient context information extraction module to predict saliency map.Meanwhile,in order to improve the accuracy and robustness of the Simple-SOD,a mosaic data augmentation method and a periodic multi-scale training strategy suitable for salient object detection are proposed.Meanwhile,the knowledge distillation method is used to train the backbone network under the classification task,which greatly improve the generalization ability of the model.Experiments show that,Simple-SOD can run up to 79 FPS on a dual-core CPU clocked at 2.4GHz,which meets the real-time requirements,while the inference speed of of other lightweight salient object detection models is about 10 FPS.(3)Model compression algorithms are used to compress the model designed in the abovementioned low computing power device scenario to further reduce the model parameters and inference time.On the one hand,the model pruning algorithm is used to prune Simple-SOD.When the pruning rate is controlled within 40%,the model accuracy loss is zero,which proves the feasibility of the pruning algorithm in the salient object detection model.On the other hand,the quantization aware training algorithm is used to quantize the Simple-SOD model,so that the quantized model can run up to 135 FPS on the CPU without loss of accuracy,which is 1.5 times of the original model.