Atrous Convolution Object Detection for Small-Object Accuracy

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing deep-learning architectures for object detection struggle with poor performance on small objects, requiring a large number of parameters and failing to achieve reasonable accuracy, especially on objects smaller than 32 pixels.

Innovation Solution

The use of atrous convolution, adaptive spatial-field fusion, and spatial pyramid pooling in an object detection model to enhance feature extraction and classification, reducing the number of parameters required while maintaining accuracy on small objects.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If existing object detection models are used to detect small objects, then detection accuracy is improved, but the number of parameters required increases significantly

Engineering Contradiction:

Improvedetection accuracyVSAvoidnumber of parameters

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The model segments the feature extraction process into multiple stages with different convolutional block configurations. Early layers use smaller kernels for local features while later layers use larger receptive fields for contextual information, allowing accurate small object detection without requiring uniformly high parameter density throughout the network.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Different regions of the feature maps are processed with different levels of detail preservation. The model applies selective processing where small object regions maintain higher resolution information while larger background regions are processed more coarsely, optimizing parameter usage based on local information requirements.

Inventive Principle:

Principle #3Local quality

2Measurement precision

If existing object detection models are used to detect small objects, then detection accuracy is improved, but computational resources increase

Engineering Contradiction:

Improvedetection accuracyVSAvoidcomputational resources

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The model applies full-resolution processing only to regions containing small objects while using downsampled processing for the remainder of the image. This partial application of high-computation operations reduces overall computational resource requirements while maintaining detection accuracy for small objects.

Inventive Principle:

Principle #16Partial or excessive action

Solution Approach 2:

The computational workload is segmented into different processing paths: a fast path for large object detection and a detailed path for small object detection. This segmentation allows the system to allocate computational resources efficiently based on the specific detection task at hand.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS12626487B2Object detection based on atrous convolution and adaptive processing

Publication Date: 2026.05.12 NORTHROP GRUMMAN SYSTEMS CORP

AI summary

Systems and methods for object detection can include obtaining one or more images and processing the one or more images with a machine-learned object detection model to generate one or more bounding boxes and one or more object classifications. The object detection model may perform atrous convolution, feature fusion, feature map generation, and prediction based on feature extraction.