Fully Connected Convolutional Neural Network for Variable Aspect Ratio Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional convolutional neural networks (CNNs) require input images to have a fixed aspect ratio, leading to errors and loss of original geometric characteristics when images with varying sizes are processed, as they cannot be trained with non-fixed fully connected layers and resizing images compromises recognition accuracy.
Innovation Solution
A fully connected convolutional neural network method that allows images of different sizes and aspect ratios to be processed by using a lookup table to resize images to reference sizes while maintaining their original aspect ratio, enabling the formation of one-dimensional feature values for fully connected operations without losing geometric characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional CNN resizes images to fixed size for processing, then the input image size constraint is satisfied, but the original geometric characteristics are lost
Solution Approach 1:
The patent segments the image processing pipeline into two distinct parts: a convolutional feature extraction stage that preserves variable image dimensions and a fully connected classification stage that requires fixed input. By introducing a flattening operation that converts spatial feature maps into one-dimensional vectors, the system allows images of different sizes to pass through the convolutional layers without resizing, thereby maintaining geometric characteristics while still enabling the fully connected layer to receive consistent input dimensions for classification.
2Reliability
If conventional CNN uses fully connected layer with fixed weights, then the network structure is fixed, but it cannot process images with different sizes
Solution Approach 1:
The patent introduces dynamic adaptability into the network by making the flattening operation dimension-agnostic. While the fully connected layer maintains fixed weights for stability, the preceding convolutional and flattening operations dynamically adapt to different input image dimensions. The flattening operation automatically adjusts the number of elements in the one-dimensional vector based on the spatial dimensions of the input image, allowing the network to process images of varying sizes without requiring fixed input dimensions or modifying the fully connected layer weights.
3Adaptability or versatility
If Darknet19 removes fully connected layer and uses only convolutional layers, then images of different sizes can be processed, but computation increases and frame rate decreases
Solution Approach 1:
The patent extracts and removes the fully connected layer from the network architecture, retaining only the convolutional layers for feature extraction. This eliminates the computational burden and fixed size constraints associated with fully connected layers while preserving the ability to process images of different sizes. The convolutional layers alone are sufficient for extracting features from variable-sized inputs, and the system achieves both adaptability and improved computational efficiency by eliminating the redundant fully connected component.
Data Source
AI summary
A method for processing image using fully connected convolutional neural network and a circuit system are provided. The method is operated using fully connected convolutional neural network (CNN) and performed by the circuit system. In the method, an image with a length, a width and an aspect ratio is obtained. A reference image closest to the input image can be obtained by querying a lookup table that records multiple reference images with various sizes to be adapted to the fully connected CNN. The input image can be resized as the closest reference image. A convolution operation is then performed onto the resized image, and a feature cube is formed after multiple operations of convolution. The feature cube is transformed to one-dimensional feature values that are configured to be inputted to a fully connected layer for fully connected operation. An output value of the fully connected CNN is generated.