The small target recognition model based on temporal-spatial-spectral features for infrared spectral imaging detection

Infrared small targets recognition is of great significance for anti-interference detection, and it is a challenging problem. Infrared spectral imaging detection can provide spectral, radiation and motion information, which is an effective way to solve the problem. The infrared spectral imaging detection system based on acousto-optic tunable filter can acquire spectral image sequences band by band, and features the capabilities of staring imaging and rapidly changing detection bands according to scenes, which is suitable for recognizing moving targets in complex and changing scenarios. However, there are some issues with the data such as spectra's mismatch in dynamic scenes and lack of grayscale correlation between adjacent frames, there is a lack of methods for processing the spectral image sequences. Therefore, we propose the small target recognition model based on temporal-spatial-spectral features, which offers a new model paradigm for target recognition based on multispectral images acquired band by band. Firstly, we recognize the critical importance of multiscale spatial features and deep features for both complex background suppression and small object extraction, so multiscale spatial feature extraction is performed to adapt to targets of varying scales. In addition, we devise a temporal-spatial-spectral features extraction network based on improved convolutional long-short-term memory cell, addressing the spectra's mismatch and difference of adjacent frames. Validated with simulated experimental data and simulation data, the result shows that with similar radiation and motion of the different targets, the accuracy is better than 90%, which can effectively distinguish between the different target.