A method of infrared multi small target detection based on MMNet

Abstract In practical scenarios, as the infrared detection range shrinks, the size of infrared small targets will dynamically increase, and commonly used infrared small target detection and tracking algorithms will not be able to continue to detect and track them stably. The MMNet (Multi task Matching Network) method is proposed to address the above questions. Firstly, a lightweight fully convolutional neural network based on codec architecture was designed to segment infrared images, achieve background suppression and target enhancement; Then, utilizing the significant features of infrared small targets to further suppress false alarms; Finally, an adaptive threshold method is used to separate small targets. In the network structure, multiple lower sampling layers are introduced to reduce computation and increase receptive fields. The testing of real infrared images shows that this algorithm outperforms typical infrared small target detection algorithms in terms of detection rate, false alarm rate, and operation time, and is suitable for infrared small target detection in complex backgrounds. For unreliable tracking results, a Kalman filter is used to predict the target position. The experimental results on the LSOTB-TIR (Large Scale Thermal Infrared Object Tracking Benchmark) infrared dataset show that the improved algorithm has better performance, with tracking accuracy and success rate improved by 5.7% and 4.2% respectively compared to MMNet.