Object detection is an important task in machine vision applications. In recent years, with tangible advances in deep neural networks, general object detection has achieved promising performance, which promises to expand the applications of this field. However, detecting small objects in high-resolution images is still a challenging task. To fill the performance gap between detecting small objects and objects at large scales, various studies have been conducted in the fields of data augmentation, feature aggregation, model evolution, etc. However, due to the inherent nature of the problem, a definitive solution has not yet been presented. In other words, in a high-resolution image, there may be both small and large-sized objects, so using some customized methods for detecting small objects may not be able to perform the detection operation well and adaptively for objects with different sizes. To address this gap, in this research, an adaptive method for efficiently detecting small objects in high-resolution images based on dynamic sizing of multi-scale patches is presented. In the proposed method, which consists of a combination of several lightweight and modular convolutional neural networks, the patch size is variable based on image features, and this architecture allows for increased speed to avoid unnecessary calculations, while also improving recognition accuracy, especially for images with objects of various dimensions. Simulation results show that the proposed method is accurate, fast, and lightweight with an accuracy of AP=37.89, processing speed of FPS=25.6, and computational load of GFLOPs=218 on the VisDrone dataset.
