Abstract: Sweet peppers are prone to malformed fruits during the growth and development process. Machine replace manual identification and removal of deformed sweet peppers, on the one hand, it can improve the quality and yield of sweet peppers; on the other hand, it can solve the current problems of high labor costs and low efficiency. In order to realize the identification of sweet pepper fruits by robots, an improved YOLO v7-tiny target detection model was proposed to distinguish between normal and abnormal growth of sweet pepper fruits. The parameter-free attention module（SimAM） was integrated into the backbone feature extraction network to enhance the feature extraction and feature integration capabilities of the model; the original loss function CIOU was replaced with Focal-EIOU loss, Focal-EIOU can speed up model convergence and reduce loss value; the SiLU activation function was used to replace the Leaky ReLU in the original network to enhance the nonlinear feature extraction ability of the model. The test results showed that the overall recognition precision, recall rate, mAP0.5 and mAP0.5-0.95 of the improved model were 99.1%, 97.8%, 98.9% and 94.5%, compared with that before improvement, it was increased by 5.4 percentage points, 4.7 percentage points, 2.4 percentage points, and 10.7 percentage points, respectively, the model weight size was 10.6 MB, and the single image detection time was 4.2 ms. Compared with YOLO v7, scaled-YOLO v4, YOLOR-CSP target detection models, the model had the same F1 score as YOLO v7. Compared with scaled-YOLO v4, YOLOR-CSP was increased by 0.7 and 0.2 percentage points, respectively, mAP0.5-0.95 was increased by 0.6 percentage points, 1.2 percentage points and 0.2 percentage points, respectively, and the weight size was only 14.2%, 10.0%, 10.0% of the above model. The model proposed achieved small size and high precision, and it was easy to deploy on the mobile terminal, providing technical support for subsequent mechanized picking and quality grading.