Li Weile ; Shi Muqing ; Hong Zhonghua

Abstract

Traditional deep learning-based object detection networks often resize imagesduring the data preprocessing stage to achieve a uniform size and scale in thefeature map. Resizing is done to facilitate model propagation and fullyconnected classification. However, resizing inevitably leads to objectdeformation and loss of valuable information in the images. This drawbackbecomes particularly pronounced for tiny objects like distribution towers withlinear shapes and few pixels. To address this issue, we propose abandoning theresizing operation. Instead, we introduce Positional-Encoding Multi-headCriss-Cross Attention. This allows the model to capture contextual informationand learn from multiple representation subspaces, effectively enriching thesemantics of distribution towers. Additionally, we enhance Spatial PyramidPooling by reshaping three pooled feature maps into a new unified one whilealso reducing the computational burden. This approach allows images ofdifferent sizes and scales to generate feature maps with uniform dimensions andcan be employed in feature map propagation. Our SCAResNet incorporates theseaforementioned improvements into the backbone network ResNet. We evaluated ourSCAResNet using the Electric Transmission and Distribution InfrastructureImagery dataset from Duke University. Without any additional tricks, weemployed various object detection models with Gaussian Receptive Field basedLabel Assignment as the baseline. When incorporating the SCAResNet into thebaseline model, we achieved a 2.1% improvement in mAPs. This demonstrates theadvantages of our SCAResNet in detecting transmission and distribution towersand its value in tiny object detection. The source code is available athttps://github.com/LisavilaLee/SCAResNet_mmdet.