A Computer vision models based on deep learning are used in a wide variety of image processing pipelines in astronomy. Due to the volume and complexity of data that will be generated, modern astronomy is already undergoing a paradigm shift. New scientific megaprojects such as Vera Rubin, SKA, or E-ELT have created engineering challenges that have made it necessary to develop new techniques and models that take into account the nature of the data they will process. As deep learning architectures are known to function in a state where there are many more parameters than training examples, this requirement is directly related to the design of the model architecture. When considering the volume of astronomical data, this type of architecture has a high energy cost in several areas. The speeding up of convolutional neural networks can be achieved through a variety of techniques and approaches. Nevertheless, reducing computational power, memory, and energy consumption while maintaining model performance always entails a trade-off. An energy-efficient object detection model is presented here for morphological classification and galaxy location. In this presentation, we will present a number of techniques, including results demonstrating that tensor and running methods applied to the architecture's backbone yield a powerful model that approaches the results of the current SOTA. However, the most significant contribution is a 40% reduction in the number of parameters and a 30% reduction in energy consumption.