Critically Compressed Quantized Convolution Neural Network based High Frame Rate and Ultra-Low Delay Fruit External Defects Detection

Abstract

High frame rate and ultra-low delay fruit external defects detection plays a key role in high-efficiency and high-quality oriented fruit products manufacture. However, current traditional computer vision based commercial solutions still lack capability of detecting most types of deceptive external defects. Although recent researches have discovered deep learning 's great potential towards defects detection, solutions with large general CNNs are too slow to adapt to high-speed factory pipelines. This paper proposes a critically compressed separable convolution network, and bit depth degression quantization to further transform the network for FPGA acceleration, which makes the implementation of CNN on High Frame Rate and Ultra-Low Delay Vision System possible. With minimal searched specialized structure, the critically compressed separable convolution network is able to handle external quality classification task with a minuscule number of parameters. By assigning degressive bit depth to different layers according to degressive bit depth importance, the customized quantization is able to compress our network more efficiently than traditional method. The proposed network consists 0.1% weight size of MobileNet (alpha = 0.25), while only a 1.54% drop of overall accuracy on validation set is observed. The hardware estimation shows the network classification unit is able to work at 0.672 ms delay with the resolution of 100*100 and up to 6 classification units parallelly.