Research On Object Detection Based On Multi-scale Feature Fusion

With the development of deep learning technology,the detection accuracy and speed of object detection technology are constantly being bettered.At present,object detection technology has been applied in various scenes of life,such as intelligent monitoring,intelligent transportation,and self-driving.However,the temporal and spatial scale change of the object is still a challenging point in detection.Therefore,this paper studies these problems from the perspective of multi-scale features.Using spatial multi-scale features to study the problem of small objects that are difficult to detect,and on this basis,the problem of the lightweight detection algorithm is studied.Finally,the effective utilization of inter-frame information in video object detection is studied via time multi-scale features.The specific works of this article are as follows:(1)Aiming at the problem that small objects are difficult to detect,a spatial multiscale feature fusion algorithm based on the super-resolution of partial regions of interest is initiated.The algorithm proposes a mothod of small object prediction branch based on the low-level feature maps to better obtain the location information of the small objects.A feature connection module is designed to enhance the feature information of the low-level feature maps.To achieve more accurate feature mapping and further enrich the feature information of small objects,a multi-scale feature fusion module based on the superresolution of the region of interest is designed.Experiments show that the algorithm improves the overall accuracy of the detection model,especially for small objects.(2)A lightweight detection algorithm based on channel dimensionality reduction and attention mechanism is proposed.The channel dimensionality reduction module is designed to compress the number of channels of the detected feature map.To avoid the loss of accuracy caused by channel compression,the attention module is designed to guide the generation of features that pay more attention to the foreground areas.At the same time,the algorithm is further optimized for the structure of the fully connected layers of the second-stage detection network.Experiments prove that this algorithm improves the detection speed and accuracy.In addition,a boost has been achieved in the detection speed by applying the algorithm to the network structure in(1).(3)To make better use of the timing features in video object detection,a video object detection algorithm based on time multi-scale feature fusion is proposed.We use optical flow to generate non-key frame features.To enhance the feature representation of non-key frames,using the idea of dynamic programming,a non-densely connected feature fusion method is designed,and feature information of different time scales is used to enrich non-key frame features.It has also been shown by experimental results that this algorithm ameliorates the detection accuracy,and it has satisfactory detection performance in scenes such as object motion blur and motion occlusion.