FUSING LARGE KERNEL AND CROSS CONVOLUTION (LKC²) FOR EFFICIENT AND ACCURATE OBJECT DETECTION

Abstract

Advancements in object recognition have surged, thanks to Convolutional Neural network (CNN), but achieving an optimal mix of extensive coverage and low processing demands remains challenging. Conventional network architectures commonly employ small convolutional kernels, such as , which are effective at capturing local features but often fail to model broader spatial dependencies required for precise localization. Expanding the structure or shifting to attention-based systems usually raises resource usage and delays processing, limiting their use in live scenarios. To address this, a combined unit that integrates Large Kernel Convolution (LKC) with Cross Convolution (CC) called the LKC² block was developed, The LKC part uses channel-separated filters to expand the view area and better grasp overall patterns, while the CC part combines details from neighboring levels to maintain alignment in space and meaning. This dual strategy allows the system to effectively identify both immediate and distant features without excessively increasing variables. The proposed LKC² module was integrated into widely used object detection architectures, including You Only Look Once version 8 (YOLOv8) and Faster Region-based Convolutional Neural Network (R-CNN), and evaluated on standard benchmark datasets such as Microsoft Common Objects in Context (MS COCO) and Pascal Visual Object Classes (VOC). The outcomes revealed boosts in average accuracy of Intersection over Union (IoU) of 0.5, around 3.6% higher on Common Objects in Context (COCO) and about 2.9% on VOC compared to baselines. Notably, this enhancement added less than 5% to operations count, highlighting its resource efficiency. Additional tests confirmed that LKC and CC each improve results separately, but together they create a stronger combined effect, enhancing range perception and refining multi-level detail integration. Overall, findings suggest that employing broad filters and inter-layer blending can successfully tackle issues in local detail capture and broad scene understanding in network-based detectors. The LKC² unit improves detection accuracy for small or densely grouped objects while maintaining the computational efficiency required for real-world deployment, making it a promising enhancement for future vision-based systems in automation, autonomous navigation, and surveillance applications.