With the increasing popularity of digital media, the need to store and transmit large amounts of visual data has also increased. Image compression techniques can reduce file sizes and bandwidth requirements while maintaining an acceptable level of image quality. Classical techniques are designed to optimize images or videos for the Human Visual System (HVS), which takes into account the characteristics of the human eye and brain in perceiving and processing visual information. However, newer compression techniques, aiming not only at HVS optimization but also at improving performance when considering tasks driven by machines, are being developed. In this context, this paper proposes an efficient approach to enhance the performance of object detection Neural Networks (NNs) in highly compressed images, using Regions Of Interest (ROIs). We evaluate the mean Average Precision (mAP) in both Faster Region-based Convolutional Neural Network (R-CNN) and DyHead Neural Network (NN), considering two different application scenarios: generic object detection and industrial supervision. In comparison with the High Efficiency Video Coding (HEVC) standard, the proposed approach allows us to reduce the bitstream up to 40% while achieving a similar accuracy in the object detection task, regardless of the considered network. These results demonstrate that high compression ratios can be achieved while maintaining good image quality for machine visual perception in task-driven systems.