Research On Infrared Weak And Small Target Detection Method Based On Deep Learning

Infrared dim and small target detection is widely used in military and civilian fields.Compared with conventional target detection tasks,it has the characteristics of small size,low signal-to-noise ratio,and complex image background,etc.The huge application prospects and the characteristics of small targets make it a hot and challenging research topic for scholars from all over the world.Based on the research and analysis of classical infrared dim and small target detection methods,and the combination with deep learning technique,this paper proposes two infrared small target detection methods,including:According to the infrared image component model and deep learning image generation technique,a deep learning-based infrared image component decomposition network model is proposed,which decomposes the infrared image into a background image and a target image.By designing a discriminator for adversarial training,the decomposition capability of the network is further improved.In order to improve the background suppression and target enhancement effects of the target image,a target enhancement module based on feature maps is proposed,and a loss function based on false alarms and missed detections is used.Experiments show that the target enhancement module can further reinforce targets and reduce false alarms such as background edges in the target image.The class activation map(CAM)is introduced into the infrared small and dim target detection,and a weakly supervised infrared small and dim target detection method is proposed.In this method,an infrared image classification network is designed to determine whether there is a small target in the input image,and based on the CAM the target location can be determined.In order to further improve the target detection performance,a simple but fast low-rank decomposition method is proposed which can well utilize the class activation map.Experiments show that using the class activation map as the prior information of the low-rank decomposition method can improve the detection ability of small targets and achieve better results than using the class activation map and the low-rank decomposition method alone.In order to verify the effectiveness and practicability of the two methods in this paper,we test our methods on the newly released land and sky background Infrared Small Aircraft Target Detection datasets(ISATD).The test results show that the method proposed in this paper perform well.Without any fine-tuning,the component decomposition network combined with the target enhancement module achieved a detection rate of 89.22% and a false detection rate of 0% on the sequences of sky scene.And the class activation map combined with the low-rank decomposition method achieved a detection rate of 90.10%and a false detection rate of 33.54%.With simple fine-tuning to suppress the effects from complex land background,the component decomposition network combined with the target enhancement module achieved a detection rate of 78.19% and a false detection rate of 2.50% on the sequences of land scene.And the class activation map combined with the low-rank decomposition method achieved a detection rate of 99.49% and a false detection rate of 22.60%.