Study on positioning and detection of crayfish body parts based on machine vision

The present work presented an effective and accurate detection method for the identification and recognition of various body parts of crayfish. This approach enabled real-time detection of crayfish heads, tails, and pincers during the crayfish processing line. To enhance the detection model's performance in terms of parameter reduction, detection precision, and speed, modifications were made to the original SSD (Single Shot MultiBox Detector) model. These modifications involved substituting the backbone feature extraction network of the SSD model with the lightweight neural network MobileNetv3, and replacing the conventional NMS (Non-Maximum Suppression) technique with Soft-NMS (Soft Non-Maximum Suppression). Initially, a total of 500 specimens belonging to various crayfish species were gathered. Following this, the crayfish image dataset was established using image augmentation techniques. The total number of 1000 images in the dataset was finally obtained and the input images were divided into training set, test set and validation set (6:2:2). In order to realize the real-time classification and detection of crayfish body parts, the improved model of the single-stage target detector SSD was used to identify the positions of the crayfish parts, including the head, tail, and pincers. Finally, in order to evaluate the detection effect of the model, the average mAP (mean of Average Precision) and average detection time of a single image were used as the precision index and speed index, respectively. To verify our improved recognition method, both of target detection algorithms (YOLOV3-MobileNetV3 and YOLOV3-Darknet53) were also used for the comparison tests. The experimental results indicated that the detection speed of SSD model with MobileNetV3 was significantly improved. The mAP of SSD-MobileNetv3 model was 95.50% and the inference speed for the image was 30 ms. The aforementioned values exhibited an increase of 7.13%, 7.8%, and 5.02% in comparison to the mean average precision (mAP) of SSD-VGG16, YOLOV3-MobileNetv3, and YOLOV3-Darknet-53, correspondingly. Furthermore, these values resulted in a reduction in the average detection time by 94 ms, 0.12 ms, and 44.2 ms, respectively. Consequently, the current detection methodology exhibited the ability to rapidly and precisely identify several body parts of crayfish in real time. This study presented a comprehensive technological framework for the utilization of deep learning algorithms in the task of identifying and localizing the body parts of crayfish.