The existing machine vision has reached a certain level in areas such as autonomous driving, facial recognition, drone navigation, industrial detection, video surveillance, etc. However, its comprehensive performance still needs to be improved in complex environments such as high-speed motion and extreme dim conditions. At present, machine vision systems face the following three challenges:

1.How to collect good visual data.

2. Silicon based semiconductor technology has reached its limit in keeping up with Moore's Law, but there are bottlenecks in chip upgrades;

The human visual system is the most important component of our perceptual system, responsible for perceiving and processing over 80% of the information obtained by the human body. Faced with such a vast amount of information, the human visual system exhibits superior performance such as low redundancy, low power consumption, and strong robustness. This is incomparable to existing machine vision systems.Therefore, re examining the human visual system and drawing on its mechanisms for perceiving and processing information can help develop higher performance machine vision. In the human visual system, the functional and structural characteristics of the retina can assist in solving the above problems:

1.The structure of the retinal surface ensures good imaging results. If the photosensitive module in the machine vision system is made into a curved surface, it can further reduce computational power consumption;

2. Human neural networks contribute to the further development of neural network algorithms. Better algorithms can establish machine vision systems with higher performance;

3. The retina serves as the endpoint of the perception module and the starting point of information processing, and only one module completes the work of two modules. If the photosensitive module and preprocessing module in the machine vision system are integrated, not only can the overall integration be improved, but also the problem of memory wall can be avoided.

research findings

Recently, Jin Mingliang's team at Qingdao University, Su Jie's team, and Ge Shuzhi's team at the National University of Singapore published a review article outlining the fabrication of a class of retina-inspired flexible neuromorphic vision sensors with "integrated sensory-memory and computing" functions.

This article starts from the human visual system and compares and discusses the differences between the human visual system and traditional machine vision systems. Due to the diversity of visual information and the large amount of data, the use of non von Neumann structured flexible neural morphological visual sensors can effectively compensate for the limitations of traditional machine vision systems based on the von Neumann architecture.Firstly, this article discusses the imitation methods of retinal information processing mechanisms and outlines the principles and circuit implementation methods of non von Neumann computing architectures.

Secondly, this article takes two-dimensional materials as an example to introduce the general standards for selecting materials for this type of sensor from the perspectives of optoelectronic materials and computational materials, and compares the performance of relevant photodetectors at present.

Once again, in simulating the surface structure of the retina, this article introduces the advantages and manufacturing methods of flexible sensor arrays. Summarized the explorations of relevant workers in chronological order. Finally, this article analyzes the current challenges faced by non von Neumann flexible neural morphological visual sensors and provides prospects for their future development.

Source: Sensor Expert Network