TLDR: Researchers have developed ICPNet, a novel deep neural network inspired by the human visual cortex, to enable machines to perceive “illusory contours” – boundaries that humans see even without clear visual cues. Unlike previous deep learning models that struggled with these visual illusions, ICPNet incorporates multi-scale feature extraction, feedforward and feedback interactions, and shape-based edge detection. Evaluated on specialized datasets like AG-MNIST and the new AG-Fashion-MNIST, ICPNet significantly outperforms state-of-the-art methods, marking a crucial step towards more human-like machine perception and bridging the gap between AI and biological vision.
Human vision is incredibly adept at perceiving contours, which are crucial for recognizing objects and distinguishing them from their backgrounds. Interestingly, our brains can even perceive “illusory contours” – boundaries that aren’t physically present but are inferred from surrounding visual cues. A classic example is the abutting grating illusion, where a clear shape emerges from the perceived edges between parallel lines that don’t actually touch. While humans effortlessly see these hidden shapes, deep neural networks (DNNs), despite their impressive performance in many computer vision tasks, have largely failed to do so. This discrepancy highlights a significant gap between current artificial intelligence and human-level perception.
A new research paper, titled A biological vision inspired framework for machine perception of abutting grating illusory contours, by Xiao Zhang, Kai-Fu Yang, Xian-Shi Zhang, Hong-Zhi You, Hong-Mei Yan, and Yong-Jie Li, introduces a novel deep network called the Illusory Contour Perception Network (ICPNet) that takes inspiration from the intricate circuits of the human visual cortex to address this challenge.
Inspired by Biology
The researchers looked to the biological visual system, particularly the visual cortex, for clues. The human visual cortex processes information through a complex interplay of feedforward (information flowing up the hierarchy) and feedback (information flowing down from higher to lower areas) connections, organized into distinct layers. This biological mechanism allows for robust perception, even when visual input is incomplete or ambiguous. ICPNet mimics this architecture to enable machines to perceive illusory contours more effectively.
How ICPNet Works
ICPNet is designed with several key components that draw directly from biological vision:
-
Multi-scale Feature Projection (MFP) Module: Just as our eyes process visual information at various scales, the MFP module extracts multi-scale representations from the input image. This ensures that the network captures both fine details and broader contextual information.
-
Feature Interaction Attention Module (FIAM): Inspired by the interaction between feedforward and feedback pathways in the visual cortex, the FIAM is introduced to boost the integration of features from different levels of the network. This allows the network to refine its understanding of contours by combining high-level contextual information with low-level visual cues.
-
Edge Fusion Module (EFM): Humans exhibit a ‘shape bias,’ meaning we tend to focus on the overall shape of an object. To instill this in the network, an edge detection task is incorporated via the EFM. This module injects shape constraints, guiding the network to concentrate on the foreground and the actual contours, whether real or illusory.
The network employs a multi-task learning framework, simultaneously optimizing for both image classification and edge detection. This joint training helps ICPNet develop a more human-like understanding of shapes and boundaries.
Experimental Success
The researchers evaluated ICPNet on existing datasets like AG-MNIST and a newly constructed, more challenging dataset called AG-Fashion-MNIST. These datasets contain images distorted with abutting grating illusions, designed to test a network’s ability to perceive these hidden contours. The results were striking: ICPNet demonstrated significantly higher sensitivity to abutting grating illusory contours compared to state-of-the-art models, showing notable improvements in accuracy across various subsets.
While a gap still remains when compared to human performance, ICPNet represents a substantial step forward. Visual explanations generated by the researchers showed that ICPNet, unlike other models, genuinely focuses on the foreground and the illusory contours, aligning more closely with human intuition.
Also Read:
- A New Hybrid AI Detection System Combines Vision Transformers with Edge Analysis for Image Verification
- BRAIN: Learning from Evolving Brain Signals for Vision Understanding
Towards Human-Level AI
This work underscores the critical role of feedback modulation, attention mechanisms, multi-scale processing, and shape bias in perceiving illusory contours. By integrating these biological principles into deep learning models, ICPNet brings artificial intelligence closer to human-level visual perception, especially in interpreting complex and ambiguous visual inputs. The insights gained from this research could also inform the design of visual models for challenging real-world tasks such as camouflaged object detection and occlusion-aware segmentation, where objects are partially hidden or blend into their surroundings.
