Long-Range Dependence Involutional Network for Logo Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Image Recognition

Background:

• Logo detection is vital for applications like trademark infringement detection and intelligent transportation.
• Deep learning, particularly Convolutional Neural Networks (CNNs), excels at feature representation but struggles with multiscale and large aspect ratio objects in logo detection.
• Existing methods face limitations in accurately detecting logos with significant variations in size and shape.

Purpose of the Study:

• To develop a novel deep learning network, the Long-Range Dependence Involutional Network (LDI-Net), to overcome limitations in current logo detection methods.
• To enhance the detection of multiscale and large aspect ratio logos.
• To improve the overall efficiency and accuracy of logo detection systems.

Main Methods:

• Introduced Long-Range Dependence Involution (LD involution), a new operator combined with self-attention, to handle large aspect ratio challenges.
• Developed a Multilevel Representation Neural Architecture Search (MRNAS) with a novel multipath topology for detecting multiscale logos.
• Implemented an Adaptive RoI Pooling Module (ARM) to address logo deformation and improve detection efficiency.

Main Results:

• The proposed LDI-Net demonstrated significant effectiveness and efficiency in logo detection tasks.
• Experiments conducted on four benchmark datasets validated the approach's performance.
• The combination of LD involution, MRNAS, and ARM successfully tackled multiscale and large aspect ratio issues.

Conclusions:

• The LDI-Net presents a robust solution for challenging logo detection scenarios, particularly those involving variations in scale and aspect ratio.
• The novel components, including LD involution and MRNAS, contribute to superior performance in computer vision tasks.
• The approach offers a promising advancement for practical applications requiring accurate and efficient logo recognition.