To address our client’s needs, we developed a state-of-the-art computer vision model that utilizes video intelligence techniques. The model efficiently processes videos of fruits and vegetables by employing noise and shadow reduction techniques such as Fourier Transformation, Normalization, Gaussian Blur, and more. Color variations and segmentation issues are tackled using gradient vectors and K means clustering techniques on pixels.

Detection of local fruits and vegetables was configured using YOLO and Darknet.
We prepared a separate dataset of fruit and vegetable images with manually annotated quality ratings ranging from 1 (fresh) to 5 (decomposed). To predict the quality of items, we employed a pre-trained ResNet 50. Our solution was embedded in the client’s preferred physical device – NVIDIA Jetson Nano. The client used a 5MP camera with 10 ATP of recording under normal lighting for inspection purposes. Jetson Nano processed videos at 30 fps and produced an output of 10 fps.