Lightweight CNN Architecture for Fast Object Action Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional artificial intelligence systems employing convolutional neural networks for identifying actions performed by objects in an environment are slow and inefficient due to the need for varying numbers of filters and redundant layers.
Innovation Solution
A lightweight Convolutional Neural Network (CNN) with constant filter numbers and residual blocks is used to analyze images, employing full pre-activation residual blocks and super-convergence for fast training, which includes convolution, max/average pooling, and global average pooling layers to determine object actions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional convolutional neural networks are used for identifying actions performed by objects, then the system can detect object actions, but the determination process becomes slow and inefficient due to varying filter numbers and redundant layers
Solution Approach 1:
The patent applies parameter changes by establishing a fixed relationship between network depth and filter numbers. Specifically, the filter numbers in convolutional layers are set based on the layer index and input image channels, eliminating the need for varying filter numbers across different configurations. This parameter standardization directly addresses the inefficiency caused by adapting filter numbers to different usage scenarios.
Solution Approach 2:
The patent introduces asymmetric filter shapes, particularly using 1×3 and 3×1 convolutional filters in alternative convolutional layers. This asymmetric design allows the network to capture spatial relationships more efficiently with fewer filters compared to traditional symmetric 3×3 filters, thereby reducing redundancy while maintaining action identification accuracy.
2Measurement precision
If more filters and redundant layers are added to improve action identification accuracy, then classification performance may improve, but the system becomes slower and more memory-intensive
Solution Approach 1:
The patent employs partial action by selectively applying different types of convolutional layers at specific network depths. Alternative convolutional layers with asymmetric filters are introduced at particular stages rather than throughout the entire network, providing just enough additional feature extraction capability where needed while avoiding unnecessary computational overhead in other regions.
Solution Approach 2:
The network is segmented into distinct functional blocks: initial convolutional layers for basic feature extraction, alternative convolutional layers with asymmetric filters for enhanced spatial relationship detection, and global average pooling layers for feature aggregation. This segmentation allows each section to perform its specialized function efficiently, maintaining accuracy while reducing overall inference time.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A system for determining an action performed within an input image includes a memory to store one or more instructions, and a processor communicatively coupled to the memory, and configured to execute the one or more instructions in the memory. The processor employs a convolutional neural network (CNN) that includes a predefined number of initial stages for extracting one or more significant features corresponding to the input image, wherein each initial stage includes a first layer, and a residual block, and wherein the first layer is selected from a group consisting of a convolution layer, a max pooling layer, and an average pooling layer. The CNN includes a final stage for classifying the extracted significant features into one or more predefined classes, wherein the final stage is formed of a global average pooling layer, and a dense layer.