Research On Dim And Small Target Detection Algorithm Based On Deep Learning

Infrared dim and small target detection is one of important research directions in the field of computer vision.Due to the characteristics of infrared thermal imaging method,it has specific detection advantages over visible-light detection such as smart security,smart medical treatment,and night pedestrian detection.Due to the weak texture information and the small-scale range,traditional target detection algorithms have higher misdetection rates and false detection rates on dim and small target detection task.Compared with them,deep learning detection method tends to have higher detection precision and stronger generalization ability,which has received widespread attention from scholars in recent years.This thesis analyzes the characteristics of infrared images and detection framework based on deep learning,summarizes the main problems faced by the infrared dim and small target detection algorithm based on deep learning: Firstly,the weak texture information makes the detection algorithm difficult to capture its category information.Secondly,blur boundary makes it difficult for the detection algorithm to accurately predict target positions.Improving the detection precision,this paper proposes some methods of the infrared dim and small target detection algorithm based on deep learning.The main work is as follows:1.An infrared dim and small target detection algorithm based on attention mechanism is proposed.In this paper,our framework introduces an attention mechanism in CascadeRCNN,recalling different intersections multiple times and optimizing the boundary prediction results,setting a small anchor box to obtain more accurate detection results.To enhance dim texture information,this paper divides the design modules of the attention mechanism into global channel attention,local channel attention and spatial attention mechanisms.The factorization machine is introduced into the global channel attention to supply channel second-order explicit information.In the local channel attention,a local pooling method is proposed,which can better capture the spatial information of the target,and introduces multi-scale convolution on spatial attention,so that the module can adapt to targets of different scales.The effectiveness of the algorithm is verified on the FLIR infrared dataset.Small target and the overall average precisions are used as the main evaluation indicators.The experimental results show that: Compared with Cascade-RCNN detection algorithm without our attention mechanism,the method proposed in this paper has an increase of only 8.06% in the network parameters,the detection precision rate has an absolute improvement of about 1.8% on small targets.2.An infrared dim and small target detection algorithm based on deep bayesian network is proposed.The multi-stage regression detection algorithm based on the attention mechanism is limited due to the improper anchor point design method.We consider to propose an anchor-free method,introducing an error ratio and distance estimation branch to improve the precision of the frame prediction.The deep bayesian network is used to connect the loss confidence layer by layer to improve the precision of the network prediction.In order to enhance weak texture information,we propose a diamond-shaped key point sampling method,using deformable convolution to extract diamond-shaped key features to strengthen the classification and prediction branch.Meanwhile,because infrared datasets often have the problem of unbalanced target scales,this paper introduces a dynamic balancing strategy from the perspective of maximum likelihood estimation to balance the optimization difficulty between different losses.To solve the problem of optimization conflicts within target itself,this paper partially fine-tuned the equalization strategy through regularization method.Finally,the effectiveness of the algorithm is verified on the FLIR infrared dataset.The experimental results show that: compared with FCOS detection algorithms,the detection precision of the proposed algorithm has an absolute improvement of 1.3% on small targets on average,and has an absolute improvement of 2.8% compared with last proposed method.