Small Object Detection via Image Pyramid CNN Patches
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods, particularly using deep convolutional neural networks (CNNs), struggle with detecting small objects effectively due to memory limitations on graphics processing units (GPUs) and often perform better on large objects than small ones.
Innovation Solution
The approach involves breaking down large images into smaller image patches at different sizes and resolutions using an image pyramid, which are then analyzed by a convolutional neural network (CNN) to enhance small object detection, employing a scale-invariant method that trains the network with multiple iterations and combines results at the original image resolution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep convolutional neural networks are used for object detection, then detection effectiveness for large objects is improved, but detection performance for small objects deteriorates and memory usage increases
Solution Approach 1:
The patent divides the large input image into multiple smaller patches or tiles, which are then processed independently by the convolutional neural network. This segmentation approach reduces the memory burden on GPU while maintaining detection accuracy across the entire image. The network processes each patch separately and results are aggregated to form the final detection output.
Solution Approach 2:
The patent introduces a new dimension by creating an image pyramid with multiple resolutions. Instead of processing a single high-resolution image that consumes excessive memory, the system processes the same image at multiple downsampled resolutions, allowing the network to detect small objects at higher resolutions while using lower resolutions to reduce memory usage. This multi-resolution approach enables effective small object detection within memory constraints.
2Measurement precision
If the input image size is increased to improve small object detection, then detection precision for small objects is improved, but memory requirements exceed GPU capacity
Solution Approach 1:
The patent segments the large high-resolution image into smaller patches that can be processed within GPU memory limits. Each patch maintains sufficient resolution for small object detection while the overall memory requirement remains manageable. The network processes these smaller units independently and combines results to achieve comprehensive detection accuracy.
Solution Approach 2:
The patent adds a resolution dimension by constructing an image pyramid with multiple scales. Small objects are detected at higher resolutions within the pyramid, while lower resolutions provide contextual information and reduce overall memory consumption. This multi-scale approach enables the system to handle high-resolution input effectively without exceeding GPU memory capacity.
3Productivity
If existing object detection methods are applied to large images, then processing speed is maintained, but detection reliability for small objects deteriorates
Solution Approach 1:
The patent introduces a multi-resolution dimension through image pyramid construction, where small objects can be reliably detected at higher resolutions within the pyramid while maintaining processing efficiency. The system processes multiple resolutions in parallel or sequential manner, ensuring that small objects are not missed while keeping overall processing time acceptable through efficient resource utilization.
Solution Approach 2:
The patent segments the image processing task into smaller patch-level operations that can be executed efficiently. This segmentation allows the network to focus computational resources on detecting small objects within each patch while maintaining overall processing speed through parallelization and efficient aggregation of results across all patches.
Data Source
AI summary
Embodiments include apparatus and methods for training and/or using a convolutional neural network. An image pyramid is calculated from an original image including at least one object of interest. The image pyramid includes a first image under analysis and a second image under analysis. Image patches are calculated at a first predetermined size relative to the first image under analysis and a second plurality of image patches having a second predetermined size relative to the second image under analysis. The convolutional neural network is trained using the image patches and subsequent images are analyzed by the convolutional neural network using similar image patches.