Object Detection Based On Dynamic Recursive Neural Network

This thesis proposes a object detector based on the dynamic recursive neural network(DRNN),which simplifies the duplicated building blocks in deep neural network.Different from forwarding through different blocks sequentially in previous networks,we demonstrate that the DRNN can achieve better performance with fewer blocks by employing block recursively.We further add a gate structure to each block,which can adaptively decide the loop times of recursive blocks to reduce the computational cost.Since the recursive networks are hard to train,we propose the Loopy Variable Batch Normalization(LVBN)to stabilize the volatile gradient.Further,we improve the LVBN to correct statistical bias caused by the gate structure.We aim to study the multi-scale receptive fields of a single convolutional neural network to detect object of varied scales.This paper presents semanticcontext fuse module,which agglomerates context and texture by hierarchical structure.More additional information and rich receptive field bring significant improvement but generate marginal time consumption.To increase the proportion of small object,we augment training samples across different scales by Controlled anchor sampling,which improves the final accuracy of ouput.Experiments show that our detector reduces the parameters and computational cost and while outperforms the original model in term of the accuracy consistently on ImageNet-1k,COCO and Wider Face.Lastly we visualize and discuss the relation between image saliency and the number of loop time.