SCAResNet: A ResNet Variant Optimized for Tiny Object Detection in Transmission and Distribution Towers

TL;DR

The paper addresses the loss of critical information in tiny object detection caused by image resizing, focusing on transmission and distribution towers. It proposes SCAResNet, a ResNet-based backbone that eliminates resizing and employs Positional-Encoding Multi-head Criss-Cross Attention to capture rich context plus SPPRCSP to unify multi-scale features for propagation, with CSP and SE attention to control compute. On the ETDII dataset, SCAResNet delivers a notable improvement in mAP (approximately 2.1 percentage points) when paired with Gaussian Receptive Field Based Label Assignment baselines, and ablation confirms the individual and synergistic benefits of the two modules. The approach enhances tiny object detection in remote-sensing imagery, enabling more reliable monitoring of power-grid infrastructure, and provides actionable code for replication.

Abstract

Traditional deep learning-based object detection networks often resize images during the data preprocessing stage to achieve a uniform size and scale in the feature map. Resizing is done to facilitate model propagation and fully connected classification. However, resizing inevitably leads to object deformation and loss of valuable information in the images. This drawback becomes particularly pronounced for tiny objects like distribution towers with linear shapes and few pixels. To address this issue, we propose abandoning the resizing operation. Instead, we introduce Positional-Encoding Multi-head Criss-Cross Attention. This allows the model to capture contextual information and learn from multiple representation subspaces, effectively enriching the semantics of distribution towers. Additionally, we enhance Spatial Pyramid Pooling by reshaping three pooled feature maps into a new unified one while also reducing the computational burden. This approach allows images of different sizes and scales to generate feature maps with uniform dimensions and can be employed in feature map propagation. Our SCAResNet incorporates these aforementioned improvements into the backbone network ResNet. We evaluated our SCAResNet using the Electric Transmission and Distribution Infrastructure Imagery dataset from Duke University. Without any additional tricks, we employed various object detection models with Gaussian Receptive Field based Label Assignment as the baseline. When incorporating the SCAResNet into the baseline model, we achieved a 2.1% improvement in mAPs. This demonstrates the advantages of our SCAResNet in detecting transmission and distribution towers and its value in tiny object detection. The source code is available at https://github.com/LisavilaLee/SCAResNet_mmdet.

Figures (6)

• Figure 1: (a) is an original picture with dimensions $5760\times 3840$ pixels. (b) and (d) represent a transmission tower and three distribution towers, respectively, extracted from (a). (c) and (e) show the resized versions of (b) and (d) obtained by resizing (a) to $800\times 800$ pixels.
• Figure 2: Overview of our SCAResNet. The boxed section represents the modifications we made to improve ResNet.
• Figure 3: Overview of Positional-Encoding Multi-head CCA Module.
• Figure 4: Overview of SPPRCSP.
• Figure 5: Precision-Recall Curves of Contrast Experiments. The dashed line represents the baseline, while the solid line represents the detector using SCAResNet. The blue, red, and orange colors correspond to Cascade R-CNN, Faster R-CNN, and FCOS, respectively.