Research On Generic Object Detection Algorithm Based On Deep Feature Fusion

Generic object detection,as one of the fundamental and challenging problems in computer vision and digital image processing,has a wide range of applications in the fields of autonomous driving,face recognition,image retrieval,and industrial detection.Traditional generic object detection algorithms mainly rely on artificially designed operators to extract features.However,the weak generalization and robustness of this features restrict the further development of traditional generic object detection algorithms.With the development of big data and computer hardware in recent years,deep learning has achieved great success in the field of generic object detection with its strong feature extraction capability.This article takes the detection of generic objects as the research object and analyzes the shortcomings of the existing generic object detection algorithms,then proposes a generic object detection algorithm based on deep feature fusion,which achieves high performance in accuracy while meets the requirement for real-time detection.The main research contents of this article can be divided into the following aspects:(1)Research and analysis on SSD algorithm.Based on the detail study on the deep learning based generic object detection algorithm SSD,it is pointed out that the SSD algorithm is not ideal in detecting objects under extreme scale variation,which is easy to miss detection and false detection for small objects.By analyzing the structure of SSD,it is found that the reason for this phenomenon is that the shallow feature map lacks semantic information for object classification while the deep feature map lacks boundary information for object detection.(2)The analysis of multi-scale object problem.By analyzing the components of data sets,it is pointed out that the difficulty in detecting multi-scale objects lies in large scale variation across object instances and high rate of small object instances in data sets.Meanwhile,the advantages and disadvantages of the existing image pyramid and feature pyramid methods are analyzed.(3)The design of adaptively dense feature pyramid network(ADFPNet)algorithm.This algorithm is developed on SSD with a new proposed ADFP module,which is consisted of two components: a dense multi scales and receptive fields block(DMSRB)and an adaptively feature calibration block(AFCB).Specifically,DMSRB block extracts rich semantic information in a dense way through dilated convolutions with different dilated rates;the AFCB block calibrate the dense features to retain features contributing more and depress features contributing less.(4)Setting up an experimental environment and conducting experiments.By comparing the deep learning frameworks,the reason of choosing Pyorch to implement the proposed algorithm is explained.This paper describes the experimental environment and analyzes three widely used data sets,Pascal VOC 2007,VOC 2012 and MS coco,from various aspects including the components and evaluation criteria.Extensive experiments have been conducted on the three data sets,and then the training hyperparameters are described in detail and the experimental results are analyzed qualitatively and quantitatively.Compared with other algorithms,the proposed algorithm has achieved good performance on the three data sets.(5)Ablation experiment and analysis.Ablation experiments are conducted to verify the effectiveness of ADFP module,ADFP_L module,adaptively feature calibration block,and more default boxes.Moreover,it is also proved that ADFPNet maintains a fast detection speed through comparative experiments.