An improved YOLOv7 method to identify epidermal defects in mature litchi

Abstract

Automatic detection of defects in mature litchi plays a vital role in the classification of fruit grades. The existing method mainly relies on manual, it is difficult to meet the needs of different varieties of litchi various types of commodity packaging, and there are problems such as low efficiency, high cost and poor quality of goods. To address the above problems, this paper proposes an improved You Only Look Once(YOLO)v7 algorithm for the automatic detection of post-harvest mature litchi epidermal defects. First, a dataset of litchi defects (black spot, fall off, crack) was constructed, in which the train and test sets had 4133 and 516; Next, A Simple Parameter-Free Attention(SimAM) mechanism is introduced into the original YOLOv7 backbone network, while GSconv is used in the neck instead of convolution, and the shallow network is used instead of the deep network for lateral linking, finally, the Mish function is used as the activation function. Experimental results show the precious and mAP of the original YOLOv7 are 87.66% and 88.98%, and those of the improved YOLOv7 are 91.56% and 93.42%, improvements of 3.9% and 4.44% . A good foundation is laid for the automated classification of ripe litchi after harvesting.