Transplantation And Optimization Of Object Detection Algorithms Based On Neural Network

Since AlexNet won the ILSVRC contest in 2012,algorithms based on neural network have replaced the dominance of traditional algorithms in computer vision.Object detection,which is an important research direction in the field of computer vision,has made great progress in recent years.Face detection and text detection,as important directions in the field of object detection,have also become research hotspots.With the rise of mobile Internet and the improvement of hardware performance in the last decade,mobile device,as an algorithm deployment platform,is increasingly important than ever before.However,the deployment of face detection and text detection based on neural network on mobile devices is still in its infancy,and there are two major difficulties,that are large storage requirement and long inference delay,which greatly restrict the application of object detection on mobile devices.To solve the above difficulties,this thesis optimized the structure of the object detection model at first,and then used a series of optimization methods such as neural network pruning,layer fusion,quantization,etc.to lighten the model both in storage occupation and calculation.At last,this thesis successfully achieved the goal to offline deploy the object detection algorithm based on neural network on Android devices.This thesis includes such achievements and innovations:(1)For face detection,this thesis constructed a lightweight face detection network named FaceModel by using deep separable convolution,which is based on MTCNN.Compared with MTCNN,FaceModel reduced computation by more than 70%,in the meantime the detection accuracy dropping by only 0.4%on FDDB face detection dataset.For text detection,this thesis constructed a lightweight text detection network named OCRModel by using MobileNet-v2 as its backbone,which infrastructure is based on Pixellink.Compared with Pixellink+VGG,OCRModel reduced computation by 98%,meanwhile the detection accuracy on self-built business card data set only dropping by 1.1%.At the same time,this thesis also using Focal Loss to improve the convergence ability of the network while training the models.(2)Under the Tensorflow framework,the proposed FaceModel and OCRModel are further optimized in terms of network structure and storage.In the term of network architecture,the structure of two models is further compressed by neural network pruning and layer fusion,after which,FaceModel resulting in a 77%reduction in parameters and a 2%reduction in detection accuracy;OCRModel resulting in a 90%reduction in parameters and a 1%reduction in detection accuracy.In the term of storage,this thesis successfully converted 32-bit floating-point model to 8-bit fixed-point model by asymmetric channel-wise quantization and mixed precision training.(3)This thesis successfully designed and developed Android offline object detection application.In detail,the total storage occupation of the face detection model and text detection model is less than 1.1MB,the speed of face detection reaches 50FPS,the total delay of text detection and recognition is less than 1 second.The test results show that the proposed FaceModel and OCRModel achieve a trade-off between detection accuracy and real-time performance,and meet the requirements in terms of storage occupation and inference delay while deploying neural network-based object detection algorithms on mobile devices.