Optical remote sensing image object detection aims to precisely locate and categorize targets such as aircraft, vehicles, and ships. Challenges arise due to the vast distances in remote sensing, leading to numerous tiny objects that are hard to characterize. Additionally, complex backgrounds and environmental factors like lighting and weather conditions reduce signal-to-noise ratios, increasing detection difficulties. Although Convolutional Neural Networks (CNNs), especially those from the YOLO family, are employed for their efficient feature extraction capabilities, they perform poorly in detecting these tiny objects. The key to realize the detection of tiny objects in optical remote sensing images is to obtain sufficiently rich multi-scale feature information and clear tiny object features. Aiming at the above problems, this paper proposes a multi-scale hierarchical residual network based optical remote sensing image tiny object detection algorithm MHRM-YOLO on the basis of YOLOv5, and designs a simple and efficient Multi-scale Hierarchical Residual tiny object feature extraction Module (MHRM). This module expands on Cross Stage Partial (CSP) module by doing more layered design and using different convolutional combinations to extract features from different layered, which allows the network to obtain richer gradient information flow and output richer feature map combinations. In addition, MHRM can be easily embedded into the existing mainstream YOLO detection algorithm backbone network, which can obtain richer sensory fields at a finer granularity level and can effectively capture the contextual information of tiny objects and retain their spatial feature information. The network structure of the MHRM-YOLO algorithm is mainly divided into three parts, namely the backbone, the neck, and the head for prediction. The backbone consists of MHRM and basic convolution module, which performs finegrained feature extraction to obtain more multi-scale information and larger sensory field; the neck part uses the conventional CSP plus Path Aggregation Network (PAN) feature pyramid structure to perform multi-scale feature fusion; and the prediction part uses the optimized localization loss function to perform computation. Since tiny object detection is sensitive to positional offsets during regression, the localization loss penalty term is further improved to enhance the algorithm's ability to perceive positional offsets. The shape penalty term of the baseline CIoU localization loss has lost its effect, in this regard, the optimized loss function retains the Euclidean distance penalty term of the centroid and adjusts it to a scalable exponential function, and improves the shape penalty term to a bounding box distance penalty term, which weakens the detection algorithm's sensitivity to positional offsets, and further improves the performance of the detection algorithm. In order to validate the effectiveness of the proposed detection algorithm, MHRM-YOLO conducts systematic experiments on the challenging optical remote sensing image tiny object detection dataset AITODv2 and the tiny pedestrian dataset TinyPerson. Systematic ablation experiments are conducted for the effects between different module combinations, the effects of the loss function, the performance difference between different backbone network modules and the portability of the algorithm, and the experimental results show that both the MHRM module and the localization loss function can improve the performance of the detection algorithm. Compared with the benchmark YOLOv5 algorithm, the average detection accuracy of MHRM-YOLO on the two datasets is improved by 5.5 degrees o and 1.8 degrees o respectively, which effectively reduces the false detection rate and the leakage rate of the detection of tiny objects in optical remote sensing images. Of course, due to the use of larger-scale feature layers for detection, the MHRMYOLO detection algorithm has an increased computational volume and a slight decrease in inference speed compared with the benchmark algorithm. The algorithm still has the problem of missed detection for relatively irregularly shaped target algorithms. In addition, the experimental results show that although the detection accuracy of the MHRM-YOLO algorithm has an advantage over the mainstream detection algorithms, the detection results are generally low, much lower than the accuracy of conventional target detection, and the algorithm still has room for further optimization.
