Neural Architecture Search For Instance Segmentation

This thesis analyzes the characteristics of instance segmentation and classic network,such as PANet,FPN and Mask RCNN.After deep study on neural architecture search(NAS),based on prior knowledges,this thesis uses NAS automatically learn to generate path aggregation instance segmentation network with cross-scale information representation,the main work is as follows:After deeply analyzing and studying the requirements of instance segmentation task for network structure design,this thesis broke manual designed paradigm of information fusion structure in the network,designing a instance segmentation network structure which can be optimized.The optimized modules in the network are named Auto FPN and Auto PAN.Meanwhile,this thesis also uses NAS redesigns the mask subnet which realizes the segmentation task.Then,according to the design of optimizable network structure,this thesis designs the search space which is suitable for the instance segmentation task and analyzes how to design the search space so that the final network can have better performance,and in the case of ensuring network performance as much as possible lightweight network.According to the different design of search space and network structure,this thesis deeply studies and analyzes the search strategy of neural architecture search.In this thesis,Auto PAN and Auto FPN’s neural architecture search problem is formula with a problem which is suitable for reinforcement learning.The objective function and reward function which are suitable for instance segmentation network structure search are proposed.At the same time,this thesis propose a new mask subnet.Using gradient-based NAS search the structure of continuous convolution branch network in mask subnet.To solve the problems of traditional gradient-based NAS method,putting the entropy into objective funcution is proposed.This thesis first conducts the experiment to search the network structure of the continuous convolutional branch network of the mask subnet on the CIFAR-10 dataset and conducts a comparative experiment to find the best fusion mode.Based on COCO dataset,this thesis conducts a comparative experiment to verify the performance of the searched mask subnet and the effect of the improvement of DARTS proposed in this thesis.Subsequently,based on COCO dataset,Auto PANet is obtained by NAS,and the segmentation effect’s metric AP(average precision)on COCO dataset is 1.1 and 4 higher than the current SOTA model PANet and Mask RCNN.Finally,the transform experiment of the searched Auto PANet structure on Cityscapes dataset was carried out,and the AP was improved with 1.2 and 5.6 respectively compared with the current SOTA model PANet and Mask RCNN.