Research on lightweight detection algorithm of infrared image based on YOLOv5

At present, infrared image can show good detection effect in complex environment because of its unique imaging characteristics, which brings a lot of convenience to people's production and life. However, because the infrared image contains a great lot of unnecessary information, it has great requirements for computing power and memory capacity, so it is challenging to satisfy the demands of real-time detection on resource-constrained portable mobile devices. To solve this problem, this paper proposes a refined lightweight algorithm built upon the YOLOv5, which changes the backbone network of the model to GhostNet network, introduces CA attention mechanism into the model to improve the detection effect of small targets, and changes the loss function of the model to Focal-EIoU loss function to improve the location effect of the boundary box and quicken the model's rate of convergence. Then replace the conventional convolution used in the sampling process of the model with the space-to-depth convolution module, in order to enhance the model's detection performance on low-resolution images and enable the extraction of more feature information. According to the experimental results, the enhanced detection algorithm is just 1% less than the initial YOLOv5 algorithm, and the model volume and floating point calculation are reduced by 31.9% and 36% respectively. This is better suited for deployments that are lightweight.